Skip to content
Bud's Wiki

Open foundation-model releases

TL;DR

Section titled “TL;DR”

Coordinated open-weights releases of foundation models — either multi-domain bundles spanning several application areas (NVIDIA’s Jan 2026 CES drop across Nemotron / Cosmos / Alpamayo / GR00T / Clara) or single-model series shipped with full release packages: base + distilled + quantized variants, training code, datasets, and reference workflows (Lightricks LTX-2, Qwen3-VL-Embedding+Reranker). The pattern is to ship enough surface area — multiple variants, multiple inference backends, sometimes a technical report — that downstream teams can adopt the model without further engineering. Useful as a tracking concept for what’s openly available as a baseline / drop-in component in each modality.

Key claims

Section titled “Key claims”
  • Multi-domain open-release bundles deliberately stack models, datasets, and reference workflows together across several application areas in one announcement, rather than releasing per-model — NVIDIA’s Jan 2026 drop covers agentic LLMs (Nemotron), physical-AI world models (Cosmos), AV reasoning VLA (Alpamayo), humanoid robotics VLA (Isaac GR00T), and biomedical (Clara) in a single announcement (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry §Nemotron, §Physical AI, §Alpamayo, §Clara).
  • Single-model open releases now routinely ship full release packages spanning base + distilled + quantized variants + auxiliary models — LTX-2 ships bf16 dev, fp8 and nvfp4 quantized variants, an 8-step CFG=1 distilled variant, a distilled LoRA, and ×2 spatial and ×2 temporal latent upscalers (LTX-2: Efficient Joint Audio-Visual Foundation Model §Method, §Results); Qwen3-TTS ships 6 checkpoints across 2 sizes covering 3 task heads (Base / CustomVoice / VoiceDesign) plus a separate acoustic tokenizer (Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation §Released Models Description); HunyuanImage 3.0 ships a base + an editing-tuned Instruct variant of an 80B-total / 13B-activated MoE alongside the full technical report (HunyuanImage 3.0 Technical Report §1, §3.1); Kimi K2.5 ships four product-level modes (Instant / Thinking / Agent / Agent Swarm) plus a separately open-sourced terminal coding agent (Kimi Code) across web / mobile app / API / CLI surfaces (Kimi K2.5: Visual Agentic Intelligence §Overview).
  • Companion models for a pipeline are released together, not separately — Qwen3-VL ships the embedding and reranker series as a pair, in 2B and 8B sizes each, designed to be used in a recall→rerank pipeline (Qwen3-VL-Embedding-8B §Highlights; Qwen3-VL-Reranker-8B §Highlights).
  • Multiple inference backends are supported on release rather than via community work-arounds — Qwen3-VL-Embedding/Reranker ship usage examples for sentence-transformers, HF Transformers, vLLM, and SGLang in the model card (Qwen3-VL-Embedding-8B §Usage; Qwen3-VL-Reranker-8B §Usage); Qwen3-TTS ships HF Transformers + vLLM-Omni day-0 offline inference, with the qwen-tts PyPI package and a Gradio demo (Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation §vLLM Usage, §Launch Local Web UI Demo); DeepSeek-OCR 2 ships HF Transformers, vLLM, SGLang and Docker Model Runner snippets on the model card at release (DeepSeek-OCR 2: Visual Causal Flow HF model card).
  • Companion datasets are an integral part of the release — NVIDIA’s bundle includes 10T language tokens, 500K robotics trajectories, 455K synthetic protein structures, and 100TB of vehicle sensor data alongside the models (NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry §Physical AI, §Alpamayo, §Clara).
  • Joint technical reports cover related model families rather than per-model papers — Qwen3-VL-Embedding and Qwen3-VL-Reranker are documented in a single arXiv report (arXiv:2601.04720) covering both series (Qwen3-VL-Embedding-8B §Citation; Qwen3-VL-Reranker-8B §Citation; Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking Abstract); Qwen3-TTS likewise ships a single tech report (arXiv:2601.15621) covering the full Base/CustomVoice/VoiceDesign × 0.6B/1.7B × 12Hz/25Hz grid (Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation §Introduction).
  • Datasets are increasingly released as standalone foundation primitives, not just companions to a model — Action100M is released as the underlying data foundation behind VL-JEPA, framed by its authors as a primitive to enable downstream VLMs, world models, and robotics policies, at 147M segments / 21.3B words / 205 GB (Action100M: A Large-scale Video Action Dataset §1, §4).
  • Open process documentation is emerging as a release primitive alongside weights and datasets — HuggingFace’s Smol Training Playbook ships a ~200-page postmortem of the SmolLM3 run including failed ablations, infrastructure incidents (FSx for Lustre failure, dataloader regression), and a full training-restart case study, alongside the SmolLM3-3B weights, the nanotron training code, and the Cosmopedia / FineWeb-Edu / FineMath / Stack-Edu / Smollm-Corpus datasets (Smol Training Playbook (GPU MODE talk on SmolLM3) §Compute budget, §Training reset case study, §Infrastructure).
  • Open foundation releases at the >10B-activated MoE scale are now within reach for image generation — HunyuanImage 3.0 is positioned as the largest open-source image generative model at release, with 80B total / 13B activated parameters in a 64-expert / 8-active MoE built on the Hunyuan-A13B LLM backbone (HunyuanImage 3.0 Technical Report §1, §3.1.1).
  • Open multimodal LLM releases now claim parity with closed frontier models across vision, coding, and agentic-search benchmarks at the time of release — Kimi K2.5 reports the best OCRBench (92.3), OmniDocBench 1.5 (88.8), InfoVQA (92.6), MathVista-mini (90.1), and BrowseComp (60.6 / 74.9 / 78.4) numbers on a benchmark table that includes GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro, DeepSeek V3.2, and Qwen3-VL-235B (Kimi K2.5: Visual Agentic Intelligence Benchmark table).
  • Open releases of single-purpose perception/document models can outperform closed frontier VLMs under matched token budgets — DeepSeek-OCR 2 reaches 0.100 overall ED on OmniDocBench v1.5 at 1120 max visual tokens, beating Gemini-3 Pro (0.115 ED at the same 1120-token budget) and Seed-1.8 (0.106 ED at 5120 tokens) (DeepSeek-OCR 2: Visual Causal Flow Table 2).
  • The closed-but-API-accessible multimodal embedding cohort that open releases position against now includes a concrete next-gen datapoint — Google’s Gemini Embedding 2 is announced as a SOTA multimodal embedding model mapping text + images + video + audio + documents into a single space, API-only via Google AI Studio and the Gemini API (Gemini Embedding 2: SOTA multimodal embedding model (Google product announcement)).

Recent contributions

Section titled “Recent contributions”
  • [2026-09-01] LightNav-0: Scaling Real2Sim2Real for Zero-Shot Generalist Navigation: LightNav-0 is Light Origins’ first open embodied-model release — model + code + technical report for a generalist monocular-RGB navigation policy that ranks first on all ten monocular head-to-heads across VLN-CE / Matterport3D / HM3D / HM3D-OVON / EVT-Bench under matched forward-RGB protocol. Trained entirely in simulation via a Real2Sim2Real data engine (2,000+ scenes → 4,000+ hours), transfers zero-shot to humanoid / quadruped / aerial / wheeled robots plus Minecraft and legacy first-person games. Sibling first-embodied-release to Mistral’s Robostral Navigate: single-camera AI navigation Robostral Navigate; INSIGHT-Bench evaluation suite and broader benchmarking to follow. Collaborators listed: LimX Dynamics (deployment) + Manycore Tech (simulation).
  • [2026-08-31] Microduck Sim 2 Real: Microduck Sim2Real demo — the real-robot showcase for Pollen Robotics’s fully open sub-1 kg humanoid stack (Apache 2.0 code, CC-BY-SA-NC 3D models).
  • [2026-08-31] Microduck RL — RL training environments for a ~800 g open-source bipedal robot: Microduck RL — Pollen Robotics open-sources the complete sim2real pipeline (MJCF from Onshape, mjlab + PPO training, BAM actuator physics + DR, ONNX export with baked-in obs normalizer, runtime hot-swap contract) for a ~800 g bipedal robot under Apache 2.0.
  • [2026-08-28] OpenDM — DM0.5: An Open-World Foundation Model for General-Purpose Embodied Intelligence: OpenDM (Dexmal, Apache-2.0 code / Gemma-licensed weights) releases the DM0.5 stack: base checkpoint plus benchmark-fine-tuned variants (LIBERO, RoboTwin2.0, VLA-Arena, SO101 pick-cube via LoRA, MEM-Robodojo-Sim ARX X5 bimanual, RoboChallenge Table30V2 collection), training + SFT + LoRA + inference scripts, dataset registration API with embodiment-specific state descriptions, and a fast inference backend built on TensorRT + Triton + PyTorch flex_attention. Third recent Chinese-lab open-VLA release positioned as a direct π0.5 challenger (community framing “pi0.5 of China”), sibling to LingBot-VA, XR-1, and RynnBrain on that axis.
  • [2026-08-26] Isaac 0.5: An Open-Weight Embodied Foundation Model: Perceptron AI releases Isaac 0.5 as an open-weight 36B sparse-MoE embodied foundation model — the first open-weight VLA from Perceptron after the closed-source Introducing Perceptron Mk1 Mk1 and partner-only Introducing Perceptron Egocentric API Egocentric API. Ships weights + inference and adaptation stack + evaluation suites + robot-specific deployment configurations for 35+ embodiment configurations (31 real hardware or hardware families, plus simulated settings, a capture rig, a gamepad layout, and one aggregated multi-platform dataset). Repo at github.com/perceptron-ai-inc/isaac. Extends the open-weight embodied FM cohort alongside XR-1, LingBot-VA/VLA, Cosmos 3, and RynnBrain 1.1.
  • [2026-08-26] Qwen3.8-Flash-Next — Hybrid Attention with QSA, Gated Residual, N-gram Embedding: Qwen releases Qwen3.8-Flash-Next as an experimental preview of the Qwen4 architecture — 125B total / 6B active MoE + 51B N-gram Embedding + 4B MTP, day-0 HF weights with Transformers / vLLM / SGLang / TokenSpeed / Docker model-runner support; official Qwen3.8-Flash (1M context by default, built-in tools) ships alongside via Qwen Cloud.
  • [2026-08-21] Lightwheel AI open-sources EgoSuite-Open100K — 100K-hour egocentric human dataset with hand and body pose: Lightwheel AI open-sources EgoSuite-Open100K — 100K h annotated egocentric human dataset on Hugging Face with commercial-training license (not research-only). 15K+ workplace tasks/scenes, hand/body pose + subtask semantics + partial wrist camera; first 10K h live, staged rollout to 100K.
  • [2026-08-20] The Multimodal Intelligence of Muse Spark 1.2: Muse Spark 1.2 (MSL) preannounces an imminent open-weights release, following Muse Glimmer into the open cohort — would move MSL’s flagship out of the closed grouping it shared with GPT-5.3-Codex and Composer 2.5 at 1.1’s launch.
  • [2026-08-20] Marin scaling laws simulate the entire training trajectory of a 67B MoE run within 1%: Marin (Stanford CRFM + Open Athena) has kicked off a 535B-parameter MoE (~5e24 model-FLOPs) frontier run under their open-development model, with the training trajectory pre-forecast from the Delphi scaling ladder; the prior 67B MoE hit its target loss within 1% and heads into long-context + RL post-training next.
  • [2026-08-19] Ornith-1.5: From Self-Scaffolding to Self-Improvement: Ornith-1.5 ships three sizes (9B dense, 35B-A3B MoE, 397B MoE) under MIT with FP8/GGUF/MLX/NVFP4 quants; the mobile-friendly 9B variant runs on iPhone/Android and matches or exceeds Gemma-4-31B / Qwen3.6-35B on multiple coding tracks. Distinguishes itself from concurrent GLM-5.2 / DeepSeek-V4-Flash-0731 releases by publishing a self-improvement training recipe (proposer + harness + rollout GRPO) rather than architecture-only details.
  • [2026-08-18] DFlash 2: Keep Drafting Parallel: Inco AI ships two open-source DFlash 2 drafters — one for Qwen3.8-27B and one for Meta’s Muse Glimmer — under the incoai/*-DFlash2 Hugging Face collection. Community drafters that beat the model creators’ own co-released drafters (Muse Glimmer’s official DFlash 1, Qwen3.8-27B’s native MTP), reaching 2.7–3.4× / 3.1–4.6× throughput over autoregressive at batch 1 in SGLang. Extends the drafter-in-the-box pattern first tracked with Muse Glimmer + DFlash 1 and 1-bit Bonsai 27B + DSpark to a community drafter release cadence.
  • [2026-08-18] Mojo🔥 is now open source!: Modular open-sources the full Mojo compiler, tooling, and standard library under Apache 2.0 + LLVM exception — an infrastructure-tier programming-language release in the shape already tracked via Quadrants — high-performance multi-platform compiler for physics simulation (Python-to-multi-GPU compiler) and Genesis World — Simulation platform for general-purpose robotics & embodied AI learning. Notable release-shape dimensions: (a) staged four-year rollout (stdlib open 2024 → kernel code → tooling → compiler now), (b) asymmetric contribution policy — stdlib contributions accepted since 2024, compiler contributions deferred to end of 2026, (c) shipped one week after Mojo 1.0 (source-stability milestone). Extends the “infrastructure release” sub-pattern from simulators/renderers/compilers-as-libraries to a general-purpose systems programming language.
  • [2026-08-17] Measuring Autonomous AI Research: First public cross-lab head-to-head of 18 frontier models (Fable 5, Opus 5/4.8, Kimi K3/K2.7, GPT-5.6 Sol/Sol Pro/Luna/Terra, GPT-5.5, Sonnet 5, Grok 4.5/4.6, Qwen3.8 Max, GLM 5.2/5.3, DeepSeek V4 Pro, Muse Spark 1.1/1.2) at matched task and budget on autonomous ML research — nanoGPT speedrun closed-gap % is a cleaner per-model capability signal than static benchmarks for the AI-for-AI-research capability axis.
  • [2026-08-14] Ego-OSCAR: Egocentric Open source Stereo CAptuRe System: Ego-OSCAR (FPV Labs) is the first capture-hardware-as-open-release filed for the egocentric-data cluster: hardware CAD, wiring, watchdog firmware, capture pipeline (recording daemon, IMU sampler, time-sync tooling), companion 550 h stereo+IMU dataset across 100+ environments, per-session stereo calibration artifacts, and two corpus-wide annotation layers (209,315 free-form action segments + per-frame 3D hand reconstructions) released together. Explicit reproducibility-first release shape: sub-USD-200 COTS-only BoM, no custom PCBs, 3D-printable enclosure — the aim is to lower the activation energy for any team that wants to collect egocentric data at scale, not to compete with the per-unit fidelity of closed platforms like Project Aria. Extends the concept from prior egocentric releases which shipped datasets (RekaDaily-10k, Egocentric-1M, EgoVerse) or closed capture headsets (Ego1) toward the rig itself as the artifact.
  • [2026-08-14] Qwen3.8 open weights release — 27B dense multimodal + 2.4T-A95B MoE (Alibaba Qwen): Qwen3.8 delivers the open-weights follow-through on the July 19 tease (Qwen3.8 preview announcement — 2.4T-parameter frontier model, open-weights soon (Alibaba Qwen)) — same-day launch of a 27B dense native-multimodal tier (Apache 2.0, 262K native / 1M YaRN context) alongside the 2.4T-A95B Max-tier MoE, both on HF + ModelScope, with Unsloth Dynamic GGUFs and Unsloth Desktop fine-tuning day-0. Distinct release-shape datapoint for the cluster: coordinated multi-tier drop (dense + Max-MoE together) rather than the more common “small open first, flagship later” cadence.
  • [2026-08-13] Gemini 3.7 Flash — 50% cheaper than 3.6, ~3 weeks between releases: Google announces Gemini 3.7 Flash three weeks after 3.6 Flash with 50% price cut and algorithmic (not scale-up) intelligence gains — closed-weights API-only, illustrating the fast incremental cadence of frontier closed releases against the slower open-report cycle.
  • [2026-08-11] FluxVLA Engine — Standardized VLA Engineering Foundation: FluxVLA Engine (LimX Dynamics) is a robotics-hardware vendor’s open-source VLA engineering framework — code, model weights, and documentation released together, distributed via Alibaba Cloud PAI + GitHub. Bundles unified configuration, modular VLM/VLA/WAM interfaces, simulator adapters (Isaac Sim, LIBERO, RoboCasa), and hardware adapters (Oli, TRON 2, UR, ALOHA). Fits the vendor-open-stack pattern this page tracks — infrastructure release adjacent to their own hardware platforms rather than a standalone model drop.
  • [2026-08-10] Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device: Muse Glimmer is Meta Superintelligence Labs’ first open-weights model — a 30B dense agentic model under Apache 2.0 on Hugging Face, distilled from the closed Muse Spark teacher and shipped with a DFlash-based speculative-decoding drafter, ~4-bit quantization to <20 GB LM weights, and day-0 partner support across llama.cpp / MLX / ExecuTorch / vLLM / SGLang / Together / Fireworks / OpenRouter. Breaks the closed-only pattern MSL held through Muse Spark, Muse Image, and Muse Video; positions MSL alongside NVIDIA (Nemotron / Cosmos), Thinking Machines (Inkling), and PrismML (1-bit Bonsai 27B) in the open cohort. No benchmark table or exact parameter details in the launch post — separate methodology report deferred.
  • [2026-08-08] AI Sapiens Open Source — ROBOTIS Docs Directory: ROBOTIS’s canonical AI Sapiens open-source directory — enumerates the K1 humanoid stack’s public repos (ai_sapiens ROS 2, cyclo_lab sim/RL, URDF/MuJoCo descriptions, three motion-side forks from NVIDIA / NV Labs, DynamixelSDK), with Apache-2.0 software / CERN-OHL-P-2.0 hardware / CC BY 4.0 docs license split and BoM + STEP files marked as planned.
  • [2026-08-07] Cyclo — ROBOTIS Open-Source Modular Physical AI Framework: ROBOTIS opens CYCLO — an Apache-2.0 modular Physical AI framework unifying data collection, imitation-learning workflows, whole-body control, sim/RL (cyclo_lab), and operations across the ROBOTIS hardware lineup (AI Sapiens, AI Worker, OpenMANIPULATOR, ROBOTIS Hand, DYNAMIXEL). Hardware-vendor infrastructure-tier release (analogue of Genesis World on the sim side); notable for an explicit private-extension slot (cyclo_supervisor, cyclo_hub) that plugs alongside the public stack — a concrete instance of the “public core + reserved private tier” release shape.
  • [2026-08-06] RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data: Reka releases RekaDaily-10k under Apache 2.0 ungated — 10,312 h of unscripted first-person household video with per-clip captions, ~1,670 h in native 4K; second Reka open-data drop after RekaCS2-10k.
  • [2026-08-03] Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand: TetherIA extends the open-source-release pattern from models/datasets to physical hardware: Aero Hand Open ships full mechanical design, firmware, and MuJoCo assets for a $314 anthropomorphic dexterous hand.
  • [2026-08-03] MiniMax H3 — Open-Weights Omni-Reference Video-Audio Generation Model Card: MiniMax H3 HF model card — MiniMax’s first video entry in an otherwise LLM-only open-weights track (M2/M2.5/M2.7/M3). Ships two CFG-distilled BF16 checkpoints (FL2VA + Ref2VA), with day-0 serving on vLLM / diffusers / ComfyUI / SGLang, per-task request scripts, and reference 768p+2K MP4s. Novel anti-packaging dimension for this cluster: the H3-Context-IR prompt-refiner and H3-Regenerate-2K 2K-upscaling stages are not open-sourced — they remain hosted API components. A hosted-API-around-open-base pattern that answers the earlier “no HF link at launch” flag on MiniMax H3 — Omni-Reference open-weights video generation model.
  • [2026-07-31] DeepSeek-V4-Flash-0731 — post-training upgrade of V4-Flash, benchmarks now beating GLM-5.2: DeepSeek-V4-Flash-0731 — a new release shape for this cluster: a post-training-only API refresh of the existing V4-Flash checkpoint (same architecture, same 158B chat size as the April 24 drop) with claimed benchmark leap over V4-Pro-Preview and (per the Slack poster) over GLM-5.2 at ~3× smaller. API-only for now; V4-Pro API and DeepSeek App/Web unchanged. Adds Responses API format + Codex compatibility as inference-surface features. Distinct from the standard new-checkpoint pattern the page has been tracking.
  • [2026-07-31] MiniMax H3 — Omni-Reference open-weights video generation model: MiniMax launches H3 — first video-generation entry in MiniMax’s open-weights lineup (which was previously LLM-heavy: M2/M2.5/M2.7/M3). Omni-Reference multi-reference conditioning, distributed via hailuoai.video and the MiniMax API. Open weights promised in the launch tweet but no HF link, tech report, or benchmark numbers disclosed at filing — re-resolve once weights ship.
  • [2026-07-30] Introducing Inkling-Small — Thinking Machines' efficient open-weights MoE reasoning model: Thinking Machines releases Inkling-Small (276B / 12B-active MoE), a distilled ¼-size sibling of Inkling that matches or exceeds it on reasoning + agentic benchmarks via on-policy distillation from Inkling as teacher plus two extra weeks of agentic-coding RL.
  • [2026-07-30] Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World-Centric Foundation GUI Agents: Qwen-UI-Agent (MAI-UI, Alibaba) technical-report landing page announces a Qwen3.5-27B-based real-world GUI foundation agent with author-reproduced baselines across 7 general + 9 agentic/deep-research benchmarks; weights not yet posted at filing time, full report linked but not yet fetched.
  • [2026-07-29] Kaolin — A PyTorch Library for Accelerating 3D Deep Learning Research: Kaolin v0.18.0 is NVIDIA’s Apache-2.0 PyTorch library of GPU-optimized 3D primitives — representation-agnostic physics (Simplicits + FreeForm), GaussianSplatModel with PLY/USD I/O, DIB-R + nvdiffrast + easy_render, SPC octrees, and a USD physics schema.
  • [2026-07-27] 'The Real Debate Isn't Open-Weight AI, It's Chinese AI': Krach Institute's Giuda: Bloomberg Tech interview with Krach Institute’s Michelle Giuda reframing the open-weights debate as a US–China competition question — companion piece to the 2026-07-24 industry coalition letter Open Weights and American AI Leadership.
  • [2026-07-27] Delphi: Scaling Laws That Extrapolate 300× Past the Fit: Delphi (Marin team) is the Pythia-equivalent open scaling suite for the 2026 recipe frontier: released artifacts include checkpoints for every run across a 3e18–1e23 FLOP range (25B parameters / 600B tokens at the top), training-mixture pipelines that deterministically reproduce Nemotron-CC + StarCoderData + ProofPile 2 from public sources, the scaling recipe as a forkable CompletedAdamHParams class, the development methodology as an add_scaling_heuristic Claude Code skill, and plot-ready per-figure data (each row carries a wandb_url). The value-add over Pythia is a falsifiable baseline recipe: other groups can fork the recipe, vary one part, and compare against the baseline at whatever scale they can afford — moving open scaling suites from “checkpoints as artifact” to “recipe + checkpoints as artifact”.
  • [2026-07-25] Open Weights and American AI Leadership: Joint industry policy letter signed by 25 orgs (NVIDIA, Meta, Microsoft, Mistral, Hugging Face, Arcee, BFL, IBM, Perplexity, Y Combinator, …) making the political case for open weights — a coalition manifesto from most of the labs that ship on this page, with a notable explicit defense of distillation as a legitimate technique. Distributed via Jensen Huang’s first-ever X post.
  • [2026-07-24] Open Dreamer — Open-source Dreamer 4 world-model implementation in JAX/Flax: Open Dreamer bundles training pipeline (this repo) + inference runtime (reactor-team/open-dreamer) + live in-browser demo — an open-code + open-demo reimplementation of Dreamer 4 in JAX/Flax on Minecraft/VPT data. Sits in the “open-code reproduction of a landmark closed release” pattern alongside x-jepa (JEPA line).
  • [2026-07-24] SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation: NVIDIA Research releases SANA-Video 2.0 at 5B and 14B scales — VBench 84.30 at 480p/13.2s on one H100, 720p/5s at 13.06s via Sol-Engine — extending the NVlabs/Sana open-source family with the hybrid-linear + AttnRes video backbone.
  • [2026-07-23] GLM-5.2-Vision-NVFP4 — bolting MoonViT onto frozen GLM-5.2 with a 49.5M projector: Community-produced multimodal fine-tune of GLM-5.2 that ships MoonViT-from-Kimi-K2.6 bolted on via a 49.5M projector, NVFP4 weights, Blackwell-only SGLang serving, and Truss deploy configs under MIT.
  • [2026-07-23] FLUX 3 — Real World Models: Towards Multimodal Flow Models as the Backbone of Visual Intelligence: FLUX 3 (BFL) launch signals open-weights availability of the multimodal backbone (image + video + audio + action) as “FLUX 3 Dev” following the early-access rollout, alongside API access for FLUX 3 Video, FLUX 3 Image, and FLUX 3 Action. BFL’s third generation of flagship foundation models, first to unify content creation with action prediction under one architecture and licensing plan.
  • [2026-07-22] Perceptron Egocentric — first embodied-reasoning offering, SOTA over Gemini 3.5 Flash and Gemini Robotics-ER 1.6 annotation pipelines: Perceptron Egocentric — closed, partner-only launch of a specialized embodied-reasoning VLM by Perceptron AI (founded by ex-Meta researchers Armen Aghajanyan and Akshat Shrivastava, follow-up to their May 2026 Mk1 flagship). Counter-position to the open-weights trend on this page: physical-AI perception layer shipped API-only, no weights, no report.
  • [2026-07-22] SANA — Efficient High-Resolution Image and Video Generation family (NVlabs/Sana, sol-engine branch): NVlabs/Sana packages seven distinct projects (SANA T2I, SANA-1.5, SANA-Sprint, SANA-Video/LongSANA, SANA-WM, SANA-Streaming, Sol-RL) under one Apache-2.0 repo — an “architectural family” release pattern where all seven share the linear-attention DiT + DC-AE backbone but each explores a different task (image → video → world → streaming editing → RL post-training). Extends the “single-model deep-variant” pattern (LTX-2, Qwen3-VL) into a multi-task-per-architecture pattern, and the sol-engine branch now bundles a diffusion-serving framework alongside the models.
  • [2026-07-22] Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing: Mage-Flow (Microsoft) — 4B native-resolution T2I + image-editing stack shipped as six checkpoints (Base / RL-aligned / Turbo × Gen / Edit) with MIT-licensed code on GitHub and a Hugging Face model collection forthcoming. Positions as a compact open alternative to Qwen-Image (20B) and FLUX.2-Klein (9B), leading GenEval at 4B while running 1024² in sub-second on A100.
  • [2026-07-21] AlayaWorld: Interactive Long-Horizon World Modeling: AlayaWorld (Alaya Lab / Shanda) — full-stack open-source interactive video world model release: ~13B video DiT weights (merged_infer.safetensors on HuggingFace AlayaLab/AlayaWorld), inference code (GitHub), and a full technical report, licensed under the LTX-2 Community License (fine-tuned from LTX-2.3). Training code + partial training data on the roadmap. Open-weights alternative to closed real-time interactive WFMs like Genie 3 and Runway GWM-1.
  • [2026-07-20] Introducing Cosmos 3 Edge: Cosmos 3 Edge extends the June 2026 Cosmos 3 open release (Super 32B + Nano 8B) to the on-device tier: 4B open weights + Cosmos 3 Edge Policy (DROID) + post-training scripts + a companion Cosmos 3 Super 4-Step distillation checkpoint (25× faster T2I / I2V inference) shipped together on Hugging Face — same open-first pattern that NVIDIA has been running across the Cosmos line, now with an edge-hardware variant.
  • [2026-07-19] RynnBrain 1.1 — Open Embodied Foundation Models (Alibaba DAMO Academy): RynnBrain 1.1 (Alibaba DAMO) ships a three-scale unified embodied brain under Apache-2.0 — 2B, 9B, and 122B-A10B sparse-MoE checkpoints on HuggingFace + ModelScope — with a 1.1 technical report PDF, 8 capability cookbooks (adding native-3D grounding and contact-point prediction beyond the 1.0 tasks), inference recipes for HF transformers + SGLang, a paired RynnBrain-VLA for real-robot deployment, and a preserved 1.0 branch with the RynnBrain-Bench evaluation dataset. Sibling in packaging to Tencent’s Hy-Embodied-RxBrain-1.0 filed 3 days earlier; extends the Alibaba Rynn* stack (RynnWorld-4D, RynnWorld-Teleop, RynnVLA-001/002, RynnEC) that now spans world model + teleop + VLM + VLA as a vertically-integrated open release.
  • [2026-07-19] Qwen3.8 preview announcement — 2.4T-parameter frontier model, open-weights soon (Alibaba Qwen): Qwen3.8 preview announcement — 2.4T-parameter model positioned as “second only to Fable 5” with open weights promised soon; second filed 2T+-total open-weights-frontier tease in July 2026 alongside Kimi K3 (2.8T), sharpening the “Chinese-lab open-flagship vs Anthropic closed-flagship” framing that both announcements share.
  • [2026-07-16] Kimi K3 — Open Frontier Intelligence (2.8T MoE with KDA + AttnRes): Kimi K3 — 2.8T-parameter open MoE (16/896 activation) announced as the first open 3T-class model; weights promised by 2026-07-27, positioned behind Claude Fable 5 / GPT 5.6 Sol but ahead of all other tested models.
  • [2026-07-16] RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination: Tencent releases Hy-Embodied-RxBrain-1.0 (6.2B unified embodied cognition model) under Apache-2.0, with technical report, inference code, and weights on HuggingFace.
  • [2026-07-15] Introducing Inkling — Thinking Machines' open-weights 975B/41B MoE multimodal reasoning model: Thinking Machines ships Inkling — a 975B/41B-active MoE + Inkling-Small 276B/12B-active preview — as an open-weights bundle: original and NVFP4 checkpoints on Hugging Face, Tinker fine-tuning at launch, and day-0 multi-backend support across Together / Fireworks / Modal / Databricks / Baseten APIs plus SGLang, vLLM, TokenSpeed, llama.cpp, and Hugging Face transformers. Distinct release-package dimension for this cluster: shipped alongside a hosted fine-tuning platform (Tinker) plus a Playground for base-model evaluation, positioning the release around customization rather than raw-benchmark leadership.
  • [2026-07-14] Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model: Xiaomi-Robotics-U0 releases a 38B AR multimodal world-foundation model (built on Qwen3-32B + EMU3.5 + IBQ) with code and checkpoints, spanning T2I, editing, multi-view embodied scene generation, embodied transfer, and video rollout under one next-token loss.
  • [2026-07-14] 1-bit Bonsai 27B — GGUF: 27B-class Reasoning at 1.125 Bits Per Weight: 1-bit Bonsai 27B (GGUF) — Apache-2.0 27B-class model at 3.9 GB deployed footprint, derived from Qwen3.6-27B with all language weights at 1.125 bpw. Ships day-0 with llama.cpp CUDA/Metal kernels, MLX Apple Silicon companion, a Ternary variant (5.9 GB / 94.6% of FP16), and a packaged DSpark speculative-decoding drafter — the “multi-variant + multi-backend at launch” packaging convention extended to a 27B-scale binary-weight model.
  • [2026-07-09] Introducing Muse Spark 1.1: Muse Spark 1.1 lands as public-preview Meta Model API — closed / API-only, no weights or tech report; the direct successor to the original Muse Spark launch, completing the MSL closed-frontier cohort alongside GPT-5.3-Codex and Composer 2.5.
  • [2026-07-09] LingBot-World 2.0 / LingBot-World-Infinity — Infinite Worlds with Versatile Interactions: LingBot-World 2.0 — fifth Robbyant/Ant Group LingBot open release (after LingBot-VLA, LingBot-World v1, LingBot-VA, LingBot-Video); 14B causal-fast checkpoint on HF + ModelScope + inference code on GitHub. Notable license narrowing: v1 was Apache-2.0, v2 is CC BY-NC-SA 4.0 non-commercial share-alike — first LingBot release to drop out of the Apache-2.0 club, likely tied to WAIC 2026 productization and third-party demo hosting deals.
  • [2026-07-08] LingBot-Video: Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence: LingBot-Video — fourth Robbyant/Ant Group LingBot open release (after LingBot-VLA, LingBot-World, LingBot-VA); positioned as the first large-scale open-source MoE video foundation model for embodied intelligence, with website, GitHub, and Hugging Face checkpoints.
  • [2026-07-08] RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation: RynnWorld-4D ships as an Apache-2.0 bundle from Alibaba DAMO — paper, code (github.com/alibaba-damo-academy/RynnWorld-4D), open weights on Hugging Face, and project page — for a 4D embodied world model + inverse-dynamics policy.
  • [2026-07-08] Robostral Navigate: single-camera AI navigation: Robostral Navigate — Mistral’s first embodied model: 8B single-RGB-camera VLN policy built in-house (not from an existing open-source VLM), 76.6% R2R-CE validation-unseen; positions Mistral alongside Qwen, NVIDIA, and Google DeepMind in the physical-AI foundation-model race.
  • [2026-07-07] Arena.ai Text-to-Image Leaderboard: Arena.ai text-to-image leaderboard snapshot (2026-07-07) — the live crowd-sourced Elo ranking of the 72-entry T2I cohort this concept tracks; anchors “MAI-Image-2 = #3” style claims across product announcements. Top-4 all proprietary; Ideogram’s open model tops the open-weights cohort at #11 (Elo 1208).
  • [2026-06-30] Agents-A1: A 35B MoE Agentic Foundation Model: InternScience releases Agents-A1, a 35B MoE open agentic foundation model with SGLang/vLLM serving recipes, 262K context, and a bundled open evaluation framework — another datapoint in the mid-2026 ~35B-MoE open-agent cluster.
  • [2026-06-29] Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack: Tencent Hunyuan publishes HyVLA-0.5, a full VLA stack technical report — the “Hy-” rebrand extends from world models (HY-World 2.0, WorldPlay) into action policies, mirroring Alibaba’s Qwen-Robot Suite as a vertically-integrated embodied stack release.
  • [2026-06-27] Hallucination in World Models is Predictable and Preventable (MMBench2): MMBench2 is a fully open-source benchmark bundling 210 tasks across 10 domains (locomotion, manipulation, navigation, arcade) with live simulators + ground-truth actions/rewards/language and a paired 350M Dreamer-4-style reference world model with three runtime hallucination predictors — distinct from prior open WFM releases (Cosmos 3, V-JEPA 2) by shipping the diagnostic stack and live envs alongside the model.
  • [2026-06-23] i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models: i1 (Princeton) ships weights + JAX training code + PyTorch inference + dataset + data-processing pipeline + 300+ ablation logs, framed as a “fully open recipe” baseline. Beats the previous best fully-open T2I model by 29.5 absolute percentage points averaged across GenEval/DPG/PRISM/CVTG/LongText — a concrete instance of the open-recipe-with-controlled-experiments release shape, distinct from the multi-domain bundle (NVIDIA) and single-model-deep-variant (LTX-2, Qwen3-VL) patterns already tracked.
  • [2026-06-22] Inside NVIDIA Halos for Robotics: A Full-Stack Functional Safety System for Physical AI: NVIDIA Halos for Robotics drops the Outside-In Safety Blueprint under Apache-2.0 (NVIDIA/halos-outside-in-safety) — an open reference architecture for functional-safety agents (SIPP/SAIM/SEI/SDM) wired to the Metropolis VSS Blueprint, with Isaac Sim closed-loop testing profile; Halos Core itself is early-access (registration) rather than open
  • [2026-06-19] MolmoMotion: Language-guided 3D motion forecasting: Ai2’s MolmoMotion ships under Apache 2.0 with model weights, the 1.16M-video MolmoMotion-1M dataset, and the PointMotionBench evaluation suite — bundling model + data + benchmark in the Molmo lineage (Molmo 2 → MolmoPoint → MolmoMotion) as a coherent open stack for language-conditioned 3D motion forecasting.
  • [2026-06-17] ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation: ABC is one of the most complete bimanual robotics open releases filed — ABC-130K dataset (3,553h / 134K episodes / 195 tasks, the largest open bimanual teleop corpus), two policy architectures (ABC-DiT 2B + ABC-VLA 4.3B) with weights + training code, ABC Sim platform (MuJoCo + Blender re-rendering, 400h sim-teleop, 20 tasks), passive-leader hardware design, abcdl distributed dataloader, and over 100 hours of physical-robot evaluation logs. Adds the “full real-robot eval logs as a released artifact” dimension to the cluster.
  • [2026-06-16] Qwen-Robot Suite launch — Qwen-RobotManip, Qwen-RobotNav, Qwen-RobotWorld + RobotClaw framework (Alibaba Qwen): Alibaba ships the Qwen-Robot Suite: three coordinated embodied models (Manip + Nav + World) plus the RobotClaw agent framework, the largest single-day embodied open release filed so far.
  • [2026-06-16] Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models: Qwen-RobotManip — VLA foundation model from the Qwen team built on Qwen-VL, with GitHub at QwenLM/Qwen-RobotManip and blog at qwen.ai; trained entirely on open-source robot + ego-human datasets without proprietary collection.
  • [2026-06-14] RoboArena rolls back evaluations after benchmark hacking observed since April (Pranav Atreya announcement): First filed instance of a benchmark maintainer publicly rolling back leaderboard entries from a major open-WFM release — RoboArena’s lead author announces benchmark hacking observed since April, with Cosmos (per Slack pointer) among the removed entries, contradicting the RoboArena leadership claim in NVIDIA’s Cosmos 3 launch materials.
  • [2026-06-11] Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models: Embodied-R1.5 open-sources a full embodied-AI stack — 8B VLM weights, the VLA variant, training code, the 15B-token data system, and EmbodiedEvalKit (20+ benchmarks, 20+ models, unified to HuggingFace Parquet) — making it one of the most complete open embodied-foundation releases to date.
  • [2026-06-10] DiffusionGemma — experimental open diffusion language model from Google Gemma: Google Gemma announces DiffusionGemma — Apache 2.0 experimental open dLLM; extends the Gemma open-release program past flagship/Gemma-4-style models to architectural experiments. Tweet-only announcement (no model card / tech report at filing).
  • [2026-06-09] MiMo-V2.5-Pro-UltraSpeed: Pushing 1T-Parameter Model Generation Speed to 1000 TPS: Xiaomi MiMo-V2.5-Pro-UltraSpeed — API-only trial release of a 1T-parameter MoE running at >1000 tok/s decode via TileRT codesign. Not open-weights (API-gated, Jun 9–Jun 23 2026 trial window, 3× price for ~10× speed); continues the Xiaomi MiMo program tracked under tweet-xiaomimimo-2025-12-16 / blog-artificialanalysis-mimo-v2-pro-hunter-alpha-2026-03.
  • [2026-06-08] Spirit AI beats Nvidia on RoboArena robotics benchmark: TNW news piece — Spirit AI’s Spirit v1.6 (Hangzhou) tops RoboArena at 1,924 over NVIDIA’s Cosmos3-Nano-Policy (1,881) and DreamZero (1,763), two days after the Cosmos 3 launch. Article catalogs a parallel Chinese open-WFM ecosystem (Manifold AI WorldScape-0.2, AgiBot GenieEnvisioner-Sim2.0-2B, DexForce, XYZ Embodied AI) leading WorldArena and WorldScore alongside an extraordinary funding pace (Spirit AI: 1.5B yuan round, fourth in three months).
  • [2026-06-08] Spirit-v1.5: A Robotic Foundation Model by Spirit AI: Spirit AI ships Spirit-v1.5, an open-weights VLA (Qwen3-VL backbone + DiT action head) with base + task-specific fine-tuned checkpoints on HF, FSDP fine-tuning code, and a packaged RoboChallenge runtime wrapper — robotics datapoint for the multi-component release pattern; #1 on RoboChallenge Table30 as of Jan 11 2026.
  • [2026-06-08] Spirit-v1.5: Clean Data Is the Enemy of Great Robot Foundation Models: Spirit AI ships Spirit v1.5 as full weights + inference + per-task RoboChallenge submission checkpoints on HF/GitHub — a Qwen3-VL backbone + DiT action-head VLA that tops RoboChallenge Table30, first open-weights model to displace π0.5. Strengthens the 2026 pattern of Chinese labs leading the open VLA frontier (alongside LingBot-VLA) while DeepMind’s Gemini Robotics and PI’s π*0.6 keep training recipes private.
  • [2026-06-08] Launching Miso One — The Most Emotive AI Voice Model: Miso Labs’ launch video for Miso One (“most emotive AI voice model”), companion to the Miso TTS 8B open-weights drop.
  • [2026-06-08] Miso TTS 8B — Highly Emotive Text-to-Speech Model (Miso Labs): Miso Labs ships Miso TTS 8B, an open-weights ~8.2B-param RVQ Transformer TTS (Llama-8B backbone + Llama-300M audio decoder over Mimi codec) inspired by Sesame CSM, with SilentCipher watermarking by default and English-only support at release.
  • [2026-06-05] Ideogram 4.0 Technical Details: Open model at the forefront of design: Ideogram’s first open-weight release — 9.3B single-stream DiT trained exclusively on structured JSON captions, shipped with Apache-2.0 inference code, a separate non-commercial weight license + paid commercial tier, day-0 ComfyUI support, and a technical blog + press release + model card. A new lab-trajectory datapoint for this page: a previously closed-API image-generation startup pivoting to open-weights, with deliberate dual licensing rather than a fully permissive release.
  • [2026-06-04] NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4: NVIDIA ships the Nemotron 3 Ultra tier (550B-total / 55B-active LatentMoE, OpenMDW-1.1) on Hugging Face, completing the Nano/Super/Ultra family with full vLLM/SGLang/TRT-LLM serving recipes.
  • [2026-06-04] Nemotron 3 Ultra — 550B/55B-active hybrid Mamba-Transformer MoE open frontier model for long-running agents (NVIDIA AI): NVIDIA ships Nemotron 3 Ultra — 550B/55B-active hybrid Mamba-Transformer LatentMoE open frontier model with weights + recipe on Hugging Face / NIM / OpenRouter.
  • [2026-06-04] Gemma 4 12B — encoder-free unified multimodal model with 35M vision embedding (Google Gemma): Google Gemma announces Gemma 4 12B — Apache-2.0 unified multimodal model targeting 16GB laptops with an encoder-free 35M-parameter vision embedding module in place of a separate vision encoder.
  • [2026-06-03] HumanNet: Scaling Human-centric Video Learning to One Million Hours: HumanNet — ~1M-hour human-centric video corpus from PKU DAGroup / Ant, released on Hugging Face (DAGroup-PKU/HumanNet) with code + project page; completes Ant’s open data → world model (LingBot-World) → VLA policy (LingBot-VLA) stack.
  • [2026-06-03] Ego1 — egocentric capture headset for Physical AI (General Intelligence Labs): GI Labs announces Ego1 — first-person capture headset co-designed with a perception stack for planet-scale robotic training data; second hardware-first egocentric-data company filed (alongside Build AI / Egocentric-1M), still pre-release with no dataset numbers disclosed.
  • [2026-06-03] Project Chrono — Open-Source Multi-Physics Simulation Engine: Project Chrono — BSD-3 multi-physics simulation library (C++ core + PyChrono Anaconda wrapper) covering rigid-body DAE, deformable PDE, granular DVI/DAE, fluid-solid, and ODE regimes with a ROS2 sensor interface for robotics simulation.
  • [2026-06-02] MAI-Thinking-1: Building a Hill-Climbing Machine: Microsoft AI releases the MAI-Thinking-1 technical report (35B-active / 1T-total MoE, trained from scratch on 30T tokens of distillation-free in-house data) with detailed pre-training, mid-training, and RL methodology; weights status not specified in the report itself.
  • [2026-06-01] EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World: Cross-lab open dataset + open code (GaTech-RL2/EgoVerse) release for robot learning — 1,362 h / 80k episodes / 240 scenes / 2,087 demonstrators with EgoDB infrastructure as a living dataset, contributed by Georgia Tech, Stanford, UCSD, ETH Zürich, MIT CSAIL, Meta Reality Labs, Mecka AI, and Scale AI. Consortium-style ownership structure rather than the single-vendor pattern that dominates the cluster.
  • [2026-06-01] Nyx — GPU path-traced renderer plugin for Genesis World: Nyx — Apache 2.0 path-traced renderer for robotics, distributed as a precompiled pip wheel (gs-nyx-plugin) with examples + docs in the public repo. NVIDIA-only (CUDA 12.9+) but the renderer-as-camera-sensor design is the most concrete recent open release for robotics-grade rendering.
  • [2026-06-01] Quadrants — high-performance multi-platform compiler for physics simulation: Quadrants — Apache 2.0 Python-to-GPU compiler forked from Taichi (June 2025) and released as an independent project under the Genesis stack. Targets CUDA + ROCm + Metal + Vulkan + CPU from one Python frontend; a concrete datapoint for what a post-Taichi successor looks like at production scale.
  • [2026-06-01] Genesis World — Simulation platform for general-purpose robotics & embodied AI learning: Apache 2.0 release of the Genesis World multi-physics simulator on PyPI, with NVIDIA + AMD + macOS + CPU support. The simulator core under the Genesis World 1.0 announcement.
  • [2026-06-01] The Role of Simulation in Scalable Robotics, Genesis World 1.0, and the Path Forward: Genesis World 1.0 — open-sourced robotics simulation stack (physics + renderer + compiler) under Apache 2.0, framed as Genesis AI’s “second release in our full-stack suite”. Notable as an infrastructure-tier open release (rather than weights) staking out the explicit-physics + path-tracing position in robotics.
  • [2026-06-01] NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI: NVIDIA’s Cosmos 3 launch — Super (32B) + Nano (8B) under OpenMDW 1.1 across build.nvidia.com / Hugging Face / GitHub / NIM, with a multi-partner inference fan-out (Baseten, CoreWeave, Azure, Nebius, Deep Infra, Classmethod). New dimension for this page: a coalition governance layer (Cosmos Coalition with Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI) that brings closed-flagship video labs into a shared open WFM substrate without forcing their downstream products open.
  • [2026-06-01] Cosmos 3: Omnimodal World Models for Physical AI: NVIDIA releases Cosmos 3 (Super + Nano + T2I/I2V/Policy-DROID post-trained variants), five large synthetic datasets (PhyxSim / RobotSim / DriveSim / SynHuman / Warehouse), and the Cosmos-HUE evaluation benchmark under the OpenMDW-1.1 license — superseding the fragmented CES-2026 Cosmos Reason/Predict/Transfer stack with a single unified open model.
  • [2026-06-01] How Cosmos 3 Helps Physical AI Think Before It Acts: NVIDIA Cosmos 3 launches at GTC Taipei (June 1 2026) under the Linux Foundation OpenMDW 1.1 license — a new single-license model-centric primitive covering weights, architecture, documentation, datasets, benchmarks AND code in one. Ships in three variants (Super for post-training, Nano deployable, Edge for real-time inference coming) across build.nvidia.com, Hugging Face (nvidia/cosmos3), GitHub, and NVIDIA NIM. Adds a new dimension to this page: a coalition announced alongside the model (Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI) as part of the release package — making model + license + coalition + multi-surface deployment a single coordinated launch.
  • [2026-06-01] MiniMax M3 announcement — first open-weights model combining coding, agentic, and natively-multimodal frontier capabilities: MiniMax announces M3 — first natively-multimodal entry in the M-series, with weights and tech report promised in ~10 days; reports 59.0% SWE-Bench Pro and 1M context via “MiniMax Sparse Attention”.
  • [2026-05-28] Introducing Paris 2.0 — Decentralized Trained Open-Weight Video Generation Model: Paris 2.0 launch (Bagel Labs, May 2026) — video-generation follow-up to Paris (Paris: A Decentralized Trained Open-Weight Diffusion Model). Positioned as an open release in continuity with v1 (which was MIT-licensed weights + code); the v2 announcement post itself does not yet repeat the license terms or link weights/code, and defers details to a forthcoming technical report.
  • [2026-05-28] Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation: Qwen-Image-Bench from Alibaba Qwen — single-purpose evaluation release shipping Q-Judger weights (Qwen3.6-27B, Apache 2.0), the 1,000-prompt evaluation dataset (HF + ModelScope mirror), scoring code, ms-swift inference backend, and the arXiv paper as one coordinated package. Adds a new category to this page: the judge model as a released artifact, alongside the more common base/instruct/embedding/reranker shapes.
  • [2026-05-26] PrismML — 1-bit and Ternary Bonsai Image 4B — local-hardware diffusion image generation: PrismML teases Bonsai Image 4B (1-bit + ternary variants) — extends the Bonsai release pattern from LLMs to diffusion image generation. Pre-launch announcement only; HF model cards, runtime forks, and any whitepaper not yet linked.
  • [2026-05-23] Gemini Embedding 2: SOTA multimodal embedding model (Google product announcement): Google announces Gemini Embedding 2 — closed multimodal embedding model spanning text, images, video, audio, and documents in one space, API-only via Google AI Studio / Gemini API. No tech report, no benchmarks in the tweet. Adds a concrete datapoint to this page’s “closed-but-API-accessible” cohort that the Open Question previously enumerated abstractly; the direct open counterpart is Qwen3-VL-Embedding/Reranker (text + image + video, no audio yet).
  • [2026-05-22] DeepSeek-OCR 2: Visual Causal Flow: DeepSeek releases DeepSeek-OCR 2 as a coordinated open package — arXiv paper (2601.20552), GitHub code, HF weights, and day-0 multi-backend serving snippets (HF Transformers, vLLM, SGLang, Docker Model Runner) in the model card. Notable: a single-purpose perception model that beats Gemini-3 Pro on OmniDocBench v1.5 at the same 1120-token budget (0.100 vs 0.115 ED), reinforcing the open-release-beats-closed-frontier-at-task-parity datapoint that K2.5 also exhibits but for a much narrower scope.
  • [2026-05-22] Kimi K2.5: Visual Agentic Intelligence: Moonshot releases Kimi K2.5 — a native-multimodal LLM continued-pretrained on ~15T vision+text tokens from K2, with four product modes (Instant / Thinking / Agent / Agent Swarm) and an open-sourced terminal coding agent (Kimi Code). Notable for two dimensions this page didn’t previously cover: (a) multi-surface release (web app + mobile app + API + open-source CLI + hosted swarm preview) as a coordinated launch package, (b) headline parity-or-better claims vs GPT-5.2 / Claude 4.5 Opus / Gemini 3 Pro on the benchmark table that ships with the release.
  • [2026-05-22] Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking: The joint Qwen3-VL-Embedding + Qwen3-VL-Reranker technical report (arXiv:2601.04720) — already filed at Luma via the two HF model-card blog forms, now filed as the underlying tech report (Jiaming posted the qwen.ai blog form). Adds detail beyond the cards: multi-stage training pipeline (s0→s3), exact video tokenization (1 FPS, 4,500-token global budget across all frames), MRL + QAT objective specifics, and per-stage / granularity / quantization ablations (RQ5–RQ7). Reinforces the joint-tech-report-for-a-series pattern.
  • [2026-05-22] HunyuanImage 3.0 Technical Report: Tencent releases HunyuanImage 3.0, framed as the largest open-source image generative model at release — an 80B-total / 13B-activated MoE built on Hunyuan-A13B with a native multimodal architecture (single 16× VAE, generalized causal attention, generalized 2D RoPE, native CoT). Ships a base checkpoint and an editing-tuned Instruct variant, full technical report (arXiv:2509.23951), code+weights on HF and GitHub, and a hosted demo at hunyuan.tencent.com/chat. A new dimension for this page: open foundation releases at the >10B-activated-MoE scale in image generation.
  • [2026-05-22] Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation: Qwen ships an open TTS family — 6 checkpoints (0.6B and 1.7B at 12Hz, three task heads: Base / CustomVoice / VoiceDesign) plus a standalone acoustic tokenizer, with HF Transformers and vLLM-Omni day-0 serving, a Gradio web demo, a joint tech report (arXiv:2601.15621), and a qwen-tts PyPI package. Same single-model-series + multi-backend packaging convention previously seen in Qwen3-VL-Embedding/Reranker.
  • [2026-05-22] Smol Training Playbook (GPU MODE talk on SmolLM3): HuggingFace ships process documentation as a first-class release artifact — a ~200-page playbook + GPU MODE talk covering the full SmolLM3 pretraining run, including a ~437K GPU-hour cost breakdown (with ablations costing more than half the main run), a tensor-parallel-seed bug that forced a 1T-token training restart, and the architecture-ablation results (GQA-4/8 beats MHA/MQA; intra-document masking from day 0; cloze-format evaluation for early ablations). A new dimension for this page: the open release now includes the failure log, not just the recipe.
  • [2026-05-22] Action100M: A Large-scale Video Action Dataset: Meta FAIR releases Action100M as an open dataset (HF + GitHub) — 1.2M HowTo100M videos, 147M dense action+caption annotations, ~1.6M GPU-hours of compute baked in. Notable: a dataset-as-foundation-primitive release, with the trained model (VL-JEPA) treated as a demonstration of utility rather than the main artifact.
  • [2026-05-22] Qwen3-VL-Embedding-8B and Qwen3-VL-Reranker-8B: Qwen3-VL ships a four-model open-weights series (embedding 2B/8B + reranker 2B/8B) with training code, joint technical report, and multi-backend inference support (HF / sentence-transformers / vLLM / SGLang). Notable: a deliberate recall→rerank pipeline pairing is shipped as a single coordinated release.
  • [2026-05-22] LTX-2: Efficient Joint Audio-Visual Foundation Model: Lightricks LTX-2 — first credible open T2AV foundation model — ships with bf16 dev, fp8/nvfp4 quantized variants, an 8-step distilled variant + LoRA, and ×2 spatial/temporal upscalers, all on Hugging Face. Example of single-model release maximizing immediate-usability surface area.
  • [2026-05-22] NVIDIA Unveils New Open Models, Data and Tools to Advance AI Across Every Industry: NVIDIA’s CES 2026 multi-domain bundle — Nemotron + Cosmos + Alpamayo + GR00T + Clara, plus 10T tokens / 500K trajectories / 455K protein structures / 100TB driving data. The reference pattern for coordinated multi-domain open releases.

Open questions

Section titled “Open questions”
  • Are multi-domain bundles (NVIDIA-style) actually more useful to downstream teams than single-model releases with deeper variant packages (LTX-2 / Qwen3-VL-style)? The two strategies optimize different things — breadth of coverage vs depth of deployment-readiness — and neither has been compared head-to-head on adoption metrics.
  • Joint technical reports for model series (Qwen3-VL-Embedding+Reranker; the Qwen3-Embedding series before it; Qwen3-TTS spanning 3 task heads × 2 sizes × 2 tokenizer rates) seem to be replacing per-checkpoint papers. Does this hurt or help reproducibility / citation tracking?
  • What’s the right “minimum viable release package” for adoption? Quantized variants and distilled checkpoints clearly matter for inference; training code matters for fine-tuning; the Smol Training Playbook adds process documentation as a candidate primitive; Kimi K2.5 adds product-surface coverage (web/app/API/CLI + open-source CLI agent). The pattern is consolidating but hasn’t been formally specified.
  • Dataset-as-release (Action100M) vs model-as-release (LTX-2, Qwen3-VL, Qwen3-TTS, HunyuanImage 3.0, K2.5) vs process-documentation-as-release (Smol Training Playbook) — when companion releases include any/all of these, which is the more useful adoption primitive? Open question; HuggingFace’s claim is that process documentation is what’s missing from most open releases.
  • Distinct from open-foundation-releases: closed-but-API-accessible releases (Veo 3, Sora 2, Gemini Embedding 2 (Gemini Embedding 2: SOTA multimodal embedding model (Google product announcement)), Seed-1.6-embedding, Cohere Rerank Multimodal v3, MiniMax-Speech, ElevenLabs, Seedream 4.0, Nano Banana, GPT-Image). The gap is what filed papers usually use as the closed-source comparison baseline — HunyuanImage 3.0 explicitly benchmarks against four of these; K2.5 benchmarks against GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro, DeepSeek V3.2, and Qwen3-VL-235B in a single 40-row table.
  • The Kimi K2.5 benchmark table reports OSWorld-Verified scores of 63.3 (K2.5) vs 8.6 (GPT-5.2) and 20.7 (Gemini 3 Pro) — much larger gaps than on any other benchmark. Either OSWorld scoring is sensitive to setup choices in ways the page doesn’t disclose, or open-source post-training has pulled ahead specifically on computer-use; either way the result deserves independent reproduction.
  • DeepSeek-OCR 2 beats Gemini-3 Pro at matched 1120-token visual budget on OmniDocBench v1.5; this is now a second datapoint (alongside K2.5’s frontier-VLM-table claims) where a fully open release beats a closed frontier model on a specific axis. Worth watching whether this generalizes to broader VLM benchmarks or stays confined to narrow perception tasks.
  • The open multimodal embedding stack (Qwen3-VL-Embedding/Reranker, 2B/8B; text + image + video) does not yet include audio. Gemini Embedding 2 claims all of text + image + video + audio + docs in one space — does adding audio + documents to the open stack require new training data + new modality tokenizers, or does the recall→rerank pipeline degrade gracefully with separate per-modality embedders concatenated?

Papers

Section titled “Papers”

2026-11

Section titled “2026-11”

2026-05

Section titled “2026-05”

2026-04

Section titled “2026-04”

2026-03

Section titled “2026-03”

2026-02

Section titled “2026-02”

2026-01

Section titled “2026-01”

2025-12

Section titled “2025-12”

2025-11

Section titled “2025-11”

2025-10

Section titled “2025-10”

2025-09

Section titled “2025-09”

2025-08

Section titled “2025-08”

2025-07

Section titled “2025-07”

2025-06

Section titled “2025-06”

2025-05

Section titled “2025-05”

2025-04

Section titled “2025-04”

2025-03

Section titled “2025-03”

2024-01

Section titled “2024-01”

Unknown date

Section titled “Unknown date”