Meta releases Muse Spark Open weights for coding and agents

Meta released Muse Spark Open weights on October 15, 2026, an 816-billion-parameter mixture-of-experts model it built for coding and agentic work. On its research blog, Meta calls Muse Spark Open “the most capable US open-weight model to date.” That is Meta’s wording, not an independent ranking.

Reuters reported the same day that Meta Platforms released an open-weight version of Muse Spark for coding and AI agents. Reuters said Meta described 816 billion total parameters, 42 billion active at a time, and a context window of up to 1 million tokens. Open weights mean those trained parameters can be downloaded and modified. The license still limits how a hosted business can offer the model.

Specs on the model card

The Hugging Face model card, dated October 15, 2026, lists Muse Spark Open as an 816B-total, 42B-active mixture-of-experts transformer with a perception encoder, from Meta Superintelligence Labs. Context length is 1 million tokens at inference. Inputs are text, image, video, and audio. Output is text only. The card lists 62 layers, a vocabulary of 202,048, 256 experts with 8 activated per token, and training that included data from more than 100 languages. Meta says performance may degrade outside the strongly supported set. That set is not named on the card.

Recommended sampling is temperature 1.0, top-p 1.0, and top-k disabled. Reasoning effort defaults to xhigh. Meta says to use high or xhigh for complex problem solving, coding, and agentic tasks, and low or medium for shorter answers.

The card says this open-weights version has additional safety post-training so the checkpoint is more robust against misuse as a standalone model without system guardrails. Meta still recommends deploying it inside a broader system with guardrails for the use case, and says results for the released checkpoint should not be assumed to apply to modified models. The research blog says the model was evaluated in line with Meta’s Superintelligence Scaling Framework and met the thresholds that framework requires for release.

Where the weights and APIs are

Meta lists API access at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weight downloads are on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

The model card says the weights ship in BF16 and NVFP4, and that a local GPU run needs at least 700GB of VRAM for NVFP4 and 2TB for BF16. Frameworks named there include SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth, and Llama.cpp. Meta’s blog also lists Hugging Face Transformers and Apple MLX, with optimizations for AMD, Apple, Intel, and NVIDIA silicon. Meta says the open-weight build ships with custom auto-tunable Helion GPU kernels in SGLang, citing 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350.

For a desktop path, Meta says this is the first model in its class to run using two Apple M5 Ultra Mac Studios and MLX. The blog says each M5 Ultra’s integrated GPU can use up to 512GB of unified memory, so two machines can hold the model, and that MLX pulls the weights from Hugging Face. Reuters separately reported the two-Mac-Studio local run and availability through some API providers. Meta says developers can further tune the weights with PyTorch’s TorchTitan.

What the license gates

The weights are under the Muse Model License, marked for Muse Spark Open and dated October 15, 2026. Meta’s blog says the limited commercial terms allow many developers to use the model for free. The license grants rights to use, copy, modify (including fine-tune), distill, merge, publish, distribute, sublicense, and sell copies, subject to conditions.

Three conditions will matter in production. The copyright notice, the license, and the usage policy must travel with copies and substantial portions, including derivatives. A product or service that uses the software or a derivative must prominently display “Built with Muse Spark Open” in the interface, or in documentation, a website, or a settings menu. “Derivatives” includes models created, trained, fine-tuned, distilled, or otherwise improved using the software or its outputs.

Hosted access is the tighter gate. If a licensee or its affiliates give third parties model outputs through an API or similar interface, and their aggregate gross revenue exceeds $20 million — or the equivalent in other currencies — in any consecutive 12 months, they must notify Meta and enter a separate written agreement before offering the software or derivatives as part of that model-as-a-service business. The clause does not apply to internal use, research, commercial activity that is not model-as-a-service, or using the model as a component of the licensee’s own product when that product is not itself a model service. Meta claims no ownership of outputs. The software is provided as is, and the licensee is responsible for derivatives.

The usage policy, also dated October 15, 2026, says Muse Spark Open is not intended for people under 18. It bars uses related to military and warfare, nuclear industries or applications, and operation of critical infrastructure, transportation technologies, or heavy machinery, plus other categories in the exhibit. Use also has to stay inside applicable law.

Read “most capable” as Meta’s claim

Meta published a scorecard against models it names, including Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3, using xhigh reasoning effort for Muse Spark Open. On that table the model leads some agentic marks Meta highlights, including JobBench at 61.8 and MCP-Atlas at 89.2. It does not lead the board. Meta’s own Artificial Analysis Intelligence Index figure is 39, behind GLM 5.3 at 45 and Kimi K3 at 44. On Terminal-Bench 2.1, Meta lists 76.4 for Muse Spark Open and 85.0 for Kimi K3. The scorecard is Meta’s evaluation. The “most capable US open-weight” line is Meta’s description of the release.

Reuters added that CEO Mark Zuckerberg has argued advanced AI should not be concentrated among a small number of companies, and that open-weight models can give companies more control over their data while still requiring significant computing resources and expertise. Reuters also noted that in August Meta had released Muse Glimmer, a smaller open-weight model.

If you are choosing between a download and an API, read the Muse Model License and usage policy on the Hugging Face model card first — that page requires you to review the conditions before the model content is available — then use Meta’s research post for the partner list. Hosted inference is the shorter path if you do not have the VRAM the card specifies. For an on-device run, Meta points to MLX on two M5 Ultra Mac Studios, and to TorchTitan if you plan to tune.