Meta Releases Muse Spark Open: 816B MoE, 1M Context, Free Commercial Use

Meta just released the weights to its strongest open model. Muse Spark Open is an 816B-parameter Mixture-of-Experts with 42B active parameters per token, a 1M-token context window, and — for the first time at this scale from a US lab — a free commercial use license. The download is on Hugging Face; hosted inference is available on seven providers; the training recipe is public.

This is the strongest open-weight model from a US lab this year, by every available metric: parameter count, active parameters per token, context length, and commercial terms. The release puts Meta in a different category than Llama 3, which was a research open-weights release with a license that scared enterprise buyers. Muse Spark Open ships with the cleaner terms builders actually want to ship on.

The Numbers

The architecture: 816B total parameters, 42B active per token. For comparison, Llama 3 70B is 70B dense; Mixtral 8x22B is 141B total with 39B active; DeepSeek-V3 is 671B total with 37B active. Muse Spark Open is the largest open-weights MoE to ship, period — bigger than DeepSeek-V3 in total parameters and a larger active budget per token than any previously open-released model. The 1M context window matches Claude Opus 5 and Qwen 3 Max and exceeds every other open-weight model shipping today.

The license: Muse Model License. Free commercial use. The standard carve-outs around safety, redistribution, and named-use limitations apply, but the core commercial shipping right is granted by default. Builders who have been waiting on a license they can show legal before building revenue-bearing products on have what they wanted.

The compute claim: Meta says Muse Spark Open runs on two M5 Ultra Mac Studios. The Hacker News thread called this into question in the first 200 comments. The realistic read is that Meta is showing the active parameter budget fits in unified memory at a specific configuration, and inference is possible on a high-end desktop — not that production-scale serving is sensible on that hardware. Real serving still wants H100s, H200s, or GB200s.

Why This Matters

Three things change today.

Builders stop waiting on the license. The Llama 3 license kept enterprise buyers on the sidelines for two years — the legal review cycle was longer than the model iteration cycle. Muse Spark Open's commercial terms are the first open release from a US frontier lab that doesn't need an enterprise legal review to ship against. For a SaaS team picking a model to fine-tune on, the question just became "does the model perform" rather than "does the license permit this."

The 1M context window is now commodity. Claude Opus 5, Qwen 3 Max, and Muse Spark Open all hit 1M. Three independent labs, three different architectures, same ceiling. The "long context" tier just collapsed to the floor. What's left is the question of which model is reliable across that context. On benchmarks, Claude Opus 5 still leads on long-context recall, but the gap is now narrow enough that context length is no longer a buying criterion.

The Meta–DeepSeek–Microsoft open-weights race is now a three-front war. DeepSeek-V3 set the bar on efficiency at 671B/37B. Microsoft's MAI-Base-2 today claims 2.8T/80B with 4–12x better efficiency. Meta just released 816B/42B with the cleanest license of the three. Alibaba and Mistral are also in this tier. The "open-weights frontier" is no longer a single-model race; it's a market structure. For buyers, that means negotiating leverage and term flexibility on every release.

The Trade-Off

Meta's release is open weights, not open training data. The recipe is public; the data is not. The Muse Spark Open model card lists the data sources Meta is willing to disclose, but the full training corpus is not published. For most commercial use, this is fine. For regulated industries — financial services, healthcare, government — the disclosure gap is still a real gap.

Also unchanged: post-training. Muse Spark Open is a base model. The fine-tunes, RLHF passes, and capability work that turns a base model into a chat model are not in this release. Builders who want a chat model need to do that work themselves or wait for Meta to release fine-tunes separately.

The Bottom Line

Muse Spark Open is the strongest open-weights release from a US lab this year. Builders shipping commercial products now have a license they can defend, an architecture that competes with DeepSeek-V3 on efficiency, and a 1M context window that matches the closed-source frontier. The Mac Studio claim is marketing. The license is the headline. Test the model on your workload before committing, but the waiting period on legal is over.