Meta 公開 816B 參數 Muse Spark Open 模型權重

Meta released the weights for Muse Spark Open on October 15, 2026, a natively multimodal open-weight model with 816B total parameters, 42B active parameters, and a 1M-token context window. Meta AI Research describes it as the most capable US open-weight model to date, built for coding and agentic work such as long-horizon tasks, complex software engineering, and multi-step reasoning.

Architecture and Training Data

Muse Spark Open is a Mixture-of-Experts Transformer paired with a Perception Encoder, according to the model card published by Meta Superintelligence Labs. The language model has 62 layers, a hidden dimension of 8192, and 256 experts with eight activated per token through top-k routing. Its attention block uses gated attention with QK-norm, 128 query heads and 16 key-value heads in a grouped-query configuration, and a head dimension of 64. The vocabulary size is 202,048 with untied input and output embeddings. The Perception Encoder carries 1.8B parameters across 50 layers, with a hidden dimension of 1536, 16 heads, a 14×14 patch size, a 4096 output dimension, and sparse attention with a factor of four, under which every fourth layer plus the final layer use full attention.

The model accepts text, image, video, and audio inputs and produces text output. According to the model card, the training data includes text, images, audio, and video drawn from public sources, internet crawls, third parties, Meta’s own products and services, and synthetically generated material, covering more than 100 languages, with curation that included cleaning, processing, deduplication, and filtering. A reasoning_effort parameter controls how much the model thinks before responding, with low, medium, high, and xhigh settings and xhigh as the default; any other value returns an error. Meta recommends temperature 1.0, top_p 1.0, and top_k disabled for most deployments, and high or xhigh reasoning effort for complex problem solving, coding, and agentic tasks.

Availability and Apple Silicon Optimization

Muse Spark Open is available through API platforms Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, and the weights are downloadable on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models. It runs on inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with optimizations for AMD, Apple, Intel, and NVIDIA silicon, and developers can further tune the model using PyTorch’s TorchTitan. Weights ship in BF16 and NVFP4 versions; running locally requires a GPU with at least 700GB of VRAM for NVFP4 or 2TB for BF16.

Meta says Muse Spark Open is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with the latest MLX framework, which pulls the weights directly from Hugging Face. Apple’s unified memory architecture lets the M5 Ultra’s integrated GPU use up to 512GB per Mac Studio, and Meta says the chip’s 1.2TB/s memory bandwidth allows two machines to hold the entire model and function as a single system with a terabyte of unified memory.

Meta said it tested agentic workflows running entirely on-device with Muse Code powered by Muse Spark Open. In one test, the setup shipped real iOS changes to Ice Cubes, an open-source Mastodon client, spanning bug fixes, new features, and a full build; in another, it operated DaVinci Resolve from start to finish, examining footage frame by frame to select standout moments and reframe them as vertical crops. The open-weight version also ships with custom, auto-tunable Helion GPU kernels in SGLang, which Meta reports deliver 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350.

Meta-Reported Benchmarks and Safety Results

The model card compares Muse Spark Open, run at xhigh reasoning effort, against Inkling, Nemotron 3 Ultra, Beam, Mistral Large 4, Qwen3.8-2.4T-A95B, GLM 5.3, and Kimi K3. Meta reports a score of 39 on the Artificial Analysis Intelligence Index; the same table lists GLM 5.3 at 45, Kimi K3 at 44, and Qwen3.8-2.4T-A95B at 40. On agentic evaluations, Meta reports 61.8 on JobBench and 89.2 on MCP-Atlas, the highest scores on those tests among the models listed in the card’s table. On knowledge and reasoning, Meta reports 26.1 on HLE-Diamond without tools, 25 on AA Omniscience, and 89.0 on GPQA Diamond, where the table lists Qwen3.8-2.4T-A95B and Kimi K3 at 93.5 each.

On coding benchmarks, Meta reports 76.4 on Terminal-Bench 2.1, where the card lists Kimi K3 at 85.0 and GLM 5.3 at 83.9, along with 60 on DeepSWE 1.1 and 58 on SWE-Atlas-QnA, tests where the card lists Kimi K3 at 68 and 66. On tool-assisted multimodal tests, Meta reports 50.0 on ZeroBench, 73.6 on SimpleVQA, 80.2 on BabyVision against the table’s 85.7 for Kimi K3, and 61.5 on VideoMMEv2 against Kimi K3’s listed 65.4.

Meta said the open-weight version received additional safety post-training and was evaluated across content safety, agentic risk, privacy, and preparedness in line with its Meta Superintelligence Scaling Framework, meeting the thresholds the framework requires for release. The model card reports that Muse Spark Open refused 99.2% of jailbreaks on Meta’s internal cyber benchmark and 94.8% aggregated across PolyGuard’s cyber subset, RedCode, and CyberBench 2024; in the same setup, the card lists Kimi K3 at 75.5% and 70.2%.

Meta reports refusal rates above 98% of harmful requests on SecureBio’s BioTIER benchmark and above 99% on an internal single-turn chemistry benchmark, plus over 97% of harmful biology requests and 98% of harmful chemistry requests on internal multi-turn agentic benchmarks. The card reports a 98.1% honesty rate on the MASK benchmark, a 0% misbehavior rate on the Agentic Misalignment benchmark, and 0.5 on StrongREJECT v2, where lower scores are better. Meta cautions that the results characterize the released checkpoint under the tested conditions and that downstream modifications could degrade safety, and it recommends deploying the model within a broader AI system with guardrails suited to the use case.

License Terms and Stated Limitations

The weights are released under the Muse Model License, dated October 15, 2026, which permits free use, copying, modification including fine-tuning, creation of derivative works, distillation, and distribution, subject to conditions. Licensees must prominently display “Built with Muse Spark Open” in the user interface of each product or service that uses the software or its derivatives, or in accompanying documentation, a website, or a settings menu.

Licensees offering artificial intelligence models as a service whose aggregate gross revenue with affiliates exceeds $20 million in any consecutive 12-month period must notify Meta and enter a separate written agreement before including the software in that business. The provision does not apply to internal use, research, or other commercial activity that is not a model service, nor to use of the software as a component powering a licensee’s own product. The license provides that Meta takes no ownership of outputs generated by the software.

The attached usage policy states the model is not intended for individuals under 18 and prohibits uses including illegal activity, malware creation, weapons development, operation of critical infrastructure, and intentional deception such as fraud, disinformation, and impersonation. Meta’s stated limitations include possible inaccurate, biased, or objectionable responses; errors in multi-step reasoning in novel scenarios; degraded performance on languages outside the strongly supported set; minor quality differences under quantized inference; and unreliable viewing of very short or low-resolution videos. Meta said a full evaluation report will be available soon.