Meta opens the weights on 816 billion parameters

Meta on Thursday released the weights of Muse Spark Open, a natively multimodal model with 816bn total parameters, 42bn of them active for each token, and a one-million-token context window, the company said in its announcement. Meta described it as "the most capable US open-weight model to date", built for coding and agentic work. That ranking is Meta's own claim.

Meta's evidence is in the benchmark table on the model card, which compares Muse Spark Open with seven other open-weight models; the figures are Meta's and have not been independently reproduced. On the Artificial Analysis Intelligence Index, the table gives Muse Spark Open 39, behind GLM 5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40, and ahead of Mistral Large 4 at 38. It leads the table on several agentic and multimodal tests, including MCP-Atlas (89.2) and JobBench (61.8), and trails on others, such as Terminal-Bench 2.1. The three models above it on the index come from Chinese developers, which is where the word "US" in Meta's description does its work. TechCrunch reported Meta's claim of superior performance against other popular open-weight models; a full evaluation report is, in Meta's words, "available soon".

The hardware claim is the unusual one. Meta said the model is the first in its class to run on a desktop, using two Apple M5 Ultra Mac Studio machines and the MLX framework; each machine offers up to 512 GB of unified memory and 1.2 TB/s of memory bandwidth, so that, in Meta's words, the pair "behave like one supercomputer with a terabyte of unified memory". On conventional hardware the model card says the model needs a GPU setup with at least 700GB of video memory in its compressed NVFP4 version and 2TB in BF16. TechCrunch, citing Engadget, put the requirement at at least 1 TB of RAM.

Distribution is wide. Meta listed the weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, with hosted access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Downloads on Hugging Face sit behind an access form that asks for name, date of birth, country and affiliation. Meta also reported kernel-level speedups of 1.03 to 3.95 times in SGLang on Nvidia and AMD graphics processors.

The licence is short and specific. The Muse Model License grants use free of charge, including the right to modify, fine-tune, distil and sell the model, on three main conditions: products built on it must display "Built with Muse Spark Open"; use must follow an attached usage policy; and any company offering it, or a derivative, as a model-as-a-service product must notify Meta and sign a separate agreement once it and its affiliates exceed $20m in gross revenue over any 12 months. That threshold does not apply to internal use, research or a company using the model to power features of its own product. The usage policy bars uses including military and nuclear applications and says the model is not intended for people under 18. Open weights here mean access to the trained parameters under those terms, not an open-source release of training data.

Meta addressed the risk of releasing weights directly. It said the open version received additional safety post-training and met the release thresholds of its Meta Superintelligence Scaling Framework. The model card reports, for example, that the model refused 99.2 per cent of jailbreak attempts on Meta's internal cyber-misuse benchmark, against 75.5 per cent for Kimi K3 in the same setup, and it recommends deploying the model inside a system with its own guardrails. The card also warns that downstream modifications could degrade safety. All of these results are Meta's.

For demonstrations Meta paired the model with Muse Code, its coding terminal, running entirely on the two Mac Studios. It said the setup shipped bug fixes, new features and a full build for Ice Cubes, an open-source Mastodon client, and drove the video editor DaVinci Resolve, reviewing footage and cropping it vertically. TechCrunch described the same demonstrations. They are vendor tests rather than independent ones.

Meta's paid frontier model was the subject of Meta's first frontier model, priced at the frontier; this release is the open counterpart. Two of the models in Meta's comparison table, Mistral Large 4 and Reflection's Beam, appeared in Two open models, and no weights yet. Whether a model that needs a terabyte of memory is open in practice will depend on how many developers can afford the hardware or prefer the hosted APIs.