Skip to content
Bud's Wiki

Introducing Muse Spark Open — The Most Capable US Open-Weight Model to Date

TL;DR

Section titled “TL;DR”

Meta Superintelligence Labs releases Muse Spark Open — an 816B-total / 42B-active natively-multimodal MoE with 1M-token context, targeted at coding and long-horizon agentic work — as open weights under the custom “Muse Model License” with limited commercial terms. The release package is unusually broad: weights on Hugging Face / SageMaker / Google Cloud Model Garden / Microsoft Foundry; day-0 inference support across SGLang, vLLM, TensorRT-LLM, HF Transformers, Apple MLX, llama.cpp, and Unsloth; auto-tunable Helion GPU kernels in SGLang (claimed 1.03–3.95× kernel-level speedup on NVIDIA + AMD, 3% end-to-end on MI350); and a first-of-its-kind “runs on your desktop” story via two Apple M5 Ultra Mac Studios with the latest MLX. Billed as “the most capable US open-weight model to date” — a positional claim against the predominantly-Chinese open-frontier cohort (Kimi K3, Qwen3.8, DeepSeek-V4) and the direct open-weight follow-through on the Muse Spark line that was closed through 1.1 and only pre-announced for 1.2.

Key claims

Section titled “Key claims”
  • Muse Spark Open is a natively multimodal 816B-total / 42B-active model with a 1M-token context window, “built for coding and agentic work” and positioned on long-horizon tasks, complex software engineering, and multi-step reasoning [§Introduction].
  • The release is the first in its class to be positioned for desktop execution: two Apple M5 Ultra Mac Studios holding the full model via unified memory (up to 512 GB per Studio, 1.2 TB/s memory bandwidth) run it under the latest MLX framework [§Optimized for Apple Silicon].
  • The weights ship under a custom “Muse Model License” with limited commercial terms (not Apache / MIT / OpenMDW) — a middle position between fully-permissive open-weights and research-only [§License].
  • Day-0 inference-engine support is unusually broad: SGLang, vLLM, TensorRT-LLM, HF Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations claimed across AMD, Apple, Intel, and NVIDIA silicon [§Available Now].
  • Custom auto-tunable Helion GPU kernels are released alongside the model in SGLang, with claimed 1.03–3.95× kernel-level speedups on NVIDIA + AMD GPUs and a 3% end-to-end speedup on AMD MI350 [§Available Now].
  • Hosted-inference partner fan-out at launch spans Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI [§Introduction].
  • Model-weight distribution surface spans Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models [§Introduction].
  • Safety release package is tied to Meta’s “Superintelligence Scaling Framework” — additional safety post-training, model card with full evaluation report “coming soon”, and in-line claims of meeting the framework’s release thresholds [§Safety and Preparedness].
  • Fine-tuning surface is PyTorch TorchTitan for the server path and MLX pulling weights directly from Hugging Face for the Apple Silicon path [§Get Started].
  • End-to-end agentic desktop demos cited: shipping real iOS changes to Ice Cubes (open-source Mastodon client) end-to-end via Muse Code powered by Muse Spark Open, and driving DaVinci Resolve for frame-by-frame footage review and vertical cropping — both on-device via the Mac Studio pair [§Optimized for Apple Silicon].

Method

Section titled “Method”

Minimal architectural disclosure in the blog post: the model is advertised as a natively multimodal 816B-total / 42B-active MoE with 1M-token context, no sparsity ratio / expert-count / routing / attention-architecture details given at the blog level. A separate “evaluation methodology report” is linked for benchmarks (not fetched here), and a Muse Spark Open model card on Hugging Face (meta-models/Muse-Spark-Open-816B-A42B) is referenced for details plus a forthcoming full evaluation report. The release-engineering story is more concrete: custom auto-tunable Helion GPU kernels for SGLang, Apple MLX framework support with sharded execution across two M5 Ultra Mac Studios using unified memory (512 GB per Studio × 2 = ~1 TB effective capacity, 1.2 TB/s per-chip memory bandwidth), and TorchTitan as the recommended fine-tuning path. Hosted inference is fanned out across seven API providers at launch.

Results

Section titled “Results”

Headline benchmarks are deferred to the external evaluation methodology report (not fetched). The blog’s quantitative claims are confined to infrastructure: Helion-kernel speedups of 1.03–3.95× at the kernel level on NVIDIA + AMD, 3% end-to-end on AMD MI350. The qualitative positioning claim is “most capable US open-weight model to date” — a US-only, open-weight-only framing that implicitly accepts Chinese open frontier models (Kimi K3 at 2.8T, Qwen3.8 at 2.4T-A95B, DeepSeek-V4-Pro) as the current absolute frontier but stakes out the leading position within the US-open subset. Headline capability demonstrations are two end-to-end agentic desktop workflows: iOS bug fixes + features + full build on Ice Cubes, and footage-review-to-vertical-crop in DaVinci Resolve, both running fully on-device on the two-Mac-Studio desktop setup.

Why it’s interesting

Section titled “Why it’s interesting”

This is the open-weights shoe-drop on the Muse Spark line — the closed-API 1.1 (Introducing Muse Spark 1.1) and the pre-announcement of an open 1.2 (The Multimodal Intelligence of Muse Spark 1.2) left an open-release flag on this page’s ## Recent contributions; today’s release resolves it. The 816B/42B ratio puts Muse Spark Open at the same architectural scale regime as DeepSeek-V4 and Kimi K3 but with the explicit “US open-weight leader” framing and a licensing position (custom Muse Model License with limited commercial terms) that is deliberately not Apache 2.0 — a middle ground between MSL’s earlier Apache-2.0 Muse Glimmer release (Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device) and the gated-weights pattern used by Isaac 0.5 (Isaac 0.5: An Open-Weight Embodied Foundation Model). The desktop-execution story via dual M5 Ultra Mac Studios + MLX is novel for this scale class — 1-bit Bonsai 27B — GGUF: 27B-class Reasoning at 1.125 Bits Per Weight got a 27B-class model to 3.9 GB via 1-bit weights, but running a 816B/42B frontier model on consumer-grade desktop hardware is a different point in the design space, and whether the Helion kernels + per-partner API fan-out actually deliver on the “fastest open-weights serving” promise will show up in community benchmarking in the coming weeks.

See also

Section titled “See also”