Meta made Muse Spark Open public this Thursday, the open-weight version of its coding and agent model Muse Spark, and called the release the most capable open-weight model ever made in the United States. The shift matters because it turns a model that until now only lived inside Meta's own products into a file any developer can download, run and adapt, including on top-end home hardware like two Mac Studios with the M5 Ultra chip.
Size and architecture
Muse Spark Open has 816 billion total parameters, of which 42 billion are active per processed token, thanks to a mixture-of-experts architecture, MoE for short, which lights up only a fraction of the network at a time and lowers inference cost. The context window reaches 1 million tokens, and the model is natively multimodal: it accepts text, image, video and audio as input and answers in text. Meta says it shares the same base as Muse Spark 1.3, the closed version the company already sells via API.
Why open the weights now
Meta says it carried out extra safety training so the model resists misuse even outside the guardrails of the company's own products, and states it cleared the thresholds of its own internal risk framework, the Meta Superintelligence Scaling Framework. The release ships with day-one support on Hugging Face, on PyTorch's TorchTitan for fine-tuning, and on the SGLang and vLLM inference engines, optimized for NVIDIA, AMD and Intel GPUs. Groq and Fireworks AI already offer hosted access, and the model also appears in the catalogs of Amazon Web Services, Google Cloud and Microsoft Foundry.
What changes in practice for developers
Anyone working on long coding and automation tasks gains an open option to refactor large codebases and run agents that need to remember extensive context without depending on a closed, per-token API. In Meta's own internal test, the model scored 61.8 on JobBench, an agentic task benchmark, more than double the result of Inkling, the open model from Thinking Machines Lab. The Muse Model License allows free use, fine-tuning and redistribution for most developers, but requires a separate agreement from anyone who offers the model as a hosted service and passes US$20 million in gross revenue over 12 months, and bans military or critical infrastructure use.
The limit still untested
The cited benchmark comes from Meta itself and has no independent reproduction published yet; the company promises a full evaluation report soon. Running the 4 bit compressed version, the NVFP4 format, still requires at least 700 GB of memory, which keeps the 'runs at home' promise limited to those with two top of the line Mac Studios or access to a GPU cluster, not an everyday laptop.
The MaxAssistant view
The fact that moves the needle is not the press release, it is the file: for the first time, a model of this size and this agentic footprint leaves Meta's closed showcase and goes to Hugging Face, competing directly with the space that open Chinese models have been occupying. Before declaring a performance win, it is worth waiting for independent evaluation; what is already confirmed is the open weights, the license and the immediate availability for anyone with the infrastructure.
Sources
Meta: Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date; PyTorch: Muse Spark Open Is Now Open Weights; Hugging Face: Muse Spark Open technical overview; SiliconANGLE: Meta releases 816B-parameter Muse Spark Open for coding and agent work
