Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device

TL;DR

Muse Glimmer is Meta Superintelligence Labs’ first open-weights model, a 30B dense agentic model released under Apache 2.0 on Hugging Face and designed to run on a single consumer GPU (24–32 GB envelope) or a Mac/PC. It is positioned for always-on local agent workflows — function calling, coding, computer use, VLM-as-judge — with distillation from Muse Spark as the teacher, a small “DFlash” speculative-decoding drafter shipped alongside the main weights, and ~4-bit quantization compressing the LM to under 20 GB while leaving room for KV cache and a perception encoder. The launch bundles Hugging Face weights, developer documentation, day-0 partner support (Ollama, LM Studio, Unsloth, llama.cpp, MLX, ExecuTorch, vLLM, SGLang, Together, Fireworks, OpenRouter), and hardware-partner optimization work with AMD/Arm/Dell/Intel/NVIDIA. No benchmark numbers, exact parameter counts beyond “30B”, or training compute are disclosed in the launch post itself; a linked methodology report is deferred.

Key claims

Muse Glimmer is a 30B model optimized for local always-on agents, small enough to run on a single consumer GPU or on a Mac/PC [§intro].
Weights are open under Apache 2.0 on Hugging Face, with developer documentation and multi-backend integrations (llama.cpp, MLX, ExecuTorch) landing in the days after launch [§intro, §Get Started].
Training is a three-phase distillation-heavy pipeline: pre-training on Muse Spark’s logits, mid-training on longer-context and agent-heavy data with reasoning traces, and post-training combining SFT with on-policy distillation and reinforcement learning across general/reasoning/coding/agentic domains [§How We Trained Muse Glimmer].
Advertised capabilities include end-to-end task completion (DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench), reliable multi-schema function calling, multi-step reasoning, failure recovery (diagnose and retry on tool errors), multimodal input via a dedicated perception encoder, scaffold compatibility (OpenClaw and others), controllable reasoning-effort levels, and training on 100+ languages [§Built for Agents].
Benchmark claims are qualitative only: “performs strongly for its size class” vs Gemma4-31B and Qwen3.6-27B; no numeric table is included in the post itself, with details deferred to a linked methodology report [§Performance].
Quantization compresses the LM to ~4-bit precision, shrinking it from over 55 GB (full precision) to under 20 GB, leaving headroom for KV cache, the perception encoder, and the speculative-decoding drafter within a 24 GB or 32 GB VRAM envelope [§Fitting the Model on Your Device].
A lightweight “drafter” companion model based on DFlash (arXiv 2602.06036) proposes token blocks that the main model verifies in parallel; the launch ships quantized drafter variants to keep memory overhead small and preserve output distribution [§Faster Generation Through Speculative Decoding].
The launch bundles Hugging Face weights, hardware-partner optimization (AMD, Arm, Dell, Intel, NVIDIA), edge frameworks (llama.cpp, ExecuTorch, MLX), scale serving (vLLM, SGLang), hosted-provider access (Together, Fireworks, OpenRouter), and PyTorch TorchTitan support for further customization [§Get Started].
The model was evaluated under Meta’s Advanced AI Scaling Framework across the safety categories required for open-weight release [§How We Trained].

Method

The blog gives a phase-level training recipe but no architecture or hyperparameter details. Pre-training uses logit distillation from Muse Spark (the flagship MSL model tracked in Introducing Muse Spark 1.1) as the teacher, on a data mix similar to Spark’s. Mid-training shifts to longer-context, agent-heavy sequences with richer reasoning traces alongside organic data. Post-training combines SFT with a mix of on-policy distillation and RL across general, reasoning, coding, and agentic domains — the same OPD + RL post-training shape now common across recent frontier open releases (Nemotron-Cascade 2, MiMo-V2-Flash, Inkling-Small). Two inference-side optimizations are baked into the release: (1) ~4-bit weight quantization compresses the 30B model from >55 GB to under 20 GB, validated as introducing minimal degradation on agentic tasks; (2) a small DFlash-based drafter model proposes multi-token blocks that the main model verifies in parallel, giving speedup with identical output distribution. Both a full-precision drafter and quantized drafter variants are shipped to trade off memory against throughput. Speed is claimed to be measured on MacBook M4-Max, M5-Max, and RTX-5090, but no numbers appear in the launch post — a separate methodology report is linked.

Results

No numeric benchmarks are reported in the launch post. The only quantitative claims are memory-side: full-precision 30B would need >55 GB; the K-Quant-17GB checkpoint fits in <20 GB LM weights, leaving room in a 24 GB or 32 GB envelope for KV cache, perception encoder, and drafter. Speed claims (“fast enough for fluid conversation and real-time agent interaction”) are qualitative, with the specific numbers deferred to the linked methodology report. The comparison set called out by name is Gemma4-31B and Qwen3.6-27B (both filed on the wiki as small-dense open releases in the same size class).

Why it’s interesting

First MSL open-weights model. This is the pivotal open release from Meta Superintelligence Labs: the entire MSL lineup up to now — Muse Spark — first model from Meta Superintelligence Labs (MSL) (Muse Spark), Introducing Muse Spark 1.1 (Spark 1.1), Introducing Muse Image: Image Generation Built for Your World (Muse Image), Muse Video preview — MSL video generation model announcement (Alexandr Wang) (Muse Video preview), Introducing Muse Spark 1.1 — has been closed / API-only. Muse Glimmer breaks that pattern: MSL now has a filed presence in the open cohort alongside Kimi K3, Nemotron 3 (Nano/Super/Ultra), Inkling / Inkling-Small, and Cosmos 3 Edge.

Explicit on-device positioning. The 24–32 GB VRAM envelope, the DFlash drafter co-released with the main model, and the multi-backend day-0 support (llama.cpp / MLX / ExecuTorch) put this in the same product slot as Introducing Cosmos 3 Edge (edge WFM), 1-bit Bonsai 27B — GGUF: 27B-class Reasoning at 1.125 Bits Per Weight (1-bit Bonsai 27B at 3.9 GB), and Gemma 4 12B — encoder-free unified multimodal model with 35M vision embedding (Google Gemma) (Gemma 4 12B laptop unified multimodal) — a distinct release shape from the frontier-scale open drops (Kimi K3, MiniMax M3) that also occupy the open-weights conversation.

Distillation-first recipe as the emerging default. Logit distillation from a stronger closed teacher → OPD + RL post-training is now the second concrete instance of this exact shape at MSL after Meta’s earlier Muse-line releases, and mirrors the Introducing Inkling-Small — Thinking Machines' efficient open-weights MoE reasoning model recipe (student from Inkling teacher via on-policy distillation), the MiMo-V2.5-Pro-UltraSpeed: Pushing 1T-Parameter Model Generation Speed to 1000 TPS MOPD pipeline, and the Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation cascade RL + multi-domain OPD recipe. On-policy distillation is emerging as the default post-training regime for the open cohort.

DFlash is a paper the wiki hasn’t yet filed. The launch mentions arXiv 2602.06036 as the drafter’s underlying technique; the wiki’s speculative-decoding cluster (Speculative Decoding: Performance or Illusion?, Speculative Speculative Decoding, SpecBundle & SpecForge v0.2: Production-Ready Speculative Decoding Models and Framework, TorchSpec: Speculative Decoding Training at Scale) does not yet cover DFlash — worth flagging for a follow-up ingest.