Most frontier AI models still come with a requirement: someone else’s datacenter. Developers can build with them through an API. The model itself the infrastructure behind it and often the data path remain outside their control. Meta is testing a direction with Muse Spark Open, an open-weight agentic model large enough to compete with leading systems while still being deployable on hardware a developer or company can own.
Released by Meta Superintelligence Labs Muse Spark Open has 816 billion parameters, 42 billion active parameters per token and a 1-million-token context window. It accepts text, images, audio and video. Is built around coding, tool use, long-running agent workflows and multi-step reasoning. The large headline number needs context: Muse Spark Open uses a Mixture-of-Experts architecture so it does not activate all 816 billion parameters every time it generates a token. Its 256 experts route each token through eight experts keeping computation closer to 42 billion parameters while the full model remains available in memory.
That architecture is also what makes Meta’s computing claim interesting although “runs on your desktop” deserves an asterisk. Meta’s showcase setup uses two Mac Studios with M5 chips each supporting up to 512GB of unified memory. Apple says the model would require than 1.6TB of memory in native BF16 precision; after NVFP4 quantization its model footprint falls to about 493GB but intermediate inference memory still prevents it from fitting comfortably on a single 512GB machine. MLX therefore distributes the workload across Macs connected through high-speed communication.
The result is more than a proof that the model can load. Apple measured Muse Spark Open at 24.99 tokens per second across two M5 Ultra Mac Studios using tensor parallelism with an 8,192-token context. Four machines reached 36.73 tokens per second. A two-machine pipeline setup generated 15.74 tokens per second. Those figures turn “AI” into something more concrete: this is still expensive workstation-class infrastructure but it is meaningfully different from requiring a hyperscale GPU cluster or sending every prompt to a hosted model.
Meta is positioning that compute around agents rather than simple chat. Muse Spark Open is designed to plan across tasks call tools, inspect results recover from failed steps and continue working. Meta demonstrated the model with Muse Code making changes to the open-source Ice Cubes iOS app, including fixes, features and a full build. Another demonstration had the system operate DaVinci Resolve inspect footage frame by frame and identify moments for video crops. Its million-token context window also gives agents room to work across codebases, documents and long task histories although context capacity should not be mistaken for perfect recall.
The benchmark picture is strong. More nuanced than Meta’s description of Muse Spark Open as the “most capable US open-weight model to date.”
| Benchmark | Muse Spark Open | Qwen3.8-2.4T-A95B | GLM 5.3 | Kimi K3 |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 39 | 40 | 45 | 44 |
| JobBench | 61.8 | 52.7 | 61.4 | 54.3 |
| MCP-Atlas | 89.2 | 84.5 | 84.2 | 82.3 |
| HLE-Diamond – no tools | 26.1 | 17.3 | 16.4 | 22.2 |
| DeepSWE 1.1 | 60 | 57 | 67 | 68 |
| Terminal-Bench 2.1 | 76.4 | 82.0 | 83.9 | 85.0 |
Meta’s published table gives it 89.2 on MCP-Atlas. 61.8 On JobBench both strong agentic results. But on the Artificial Analysis Intelligence Index Muse Spark Open scores 39 behind GLM 5.3 at 45 Kimi K3 at 44 and Qwen3.8-2.4T-A95B at 40. Kimi and GLM also lead it on coding evaluations such as Terminal-Bench and DeepSWE. Independent Artificial Analysis testing currently also gives Muse Spark Open an Intelligence Index score of 39. In words it belongs in the leading open-weight tier, particularly for tool-heavy agent work but the evidence does not show it leading every global benchmark.
There is another distinction developers should notice: open-weight does not mean open source. Meta released the model under its Muse Model License, which allows use, modification, fine-tuning, derivative models and commercial deployment under stated conditions. Products using the model must prominently display “Built with Muse Spark Open.” Companies offering models directly as a Model-as-a-Service business and generating than $20 million in aggregate gross revenue over any consecutive 12-month period must notify Meta and reach a separate agreement before offering Muse Spark Open or its derivatives as part of that business. That restriction does not apply in the way to internal use or to using the model as a component of a non-MaaS product.
Meta has also tried to make the release usable beyond one hardware configuration. Muse Spark Open launches with support across SGLang, vLLM TensorRT-LLM Hugging Face Transformers, MLX, llama.cpp and Unsloth with deployment support spanning NVIDIA, AMD, Intel and Apple hardware. Hosted access is available from providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI while model catalogs include Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Local control does not remove the AI limitations. Meta says Muse Spark Open can still produce outputs make mistakes during unfamiliar multi-step tasks and show quality differences after quantization. The company recommends guardrails and human confirmation before irreversible agent actions. Its published safety results are encouraging. Several are Meta-run evaluations and the company explicitly warns that fine-tuning or other downstream modifications can change the model’s safety profile.
Muse Spark Open therefore matters for more, than its 816B parameter count. The larger shift is where powerful AI agents can live. APIs made frontier models easy to access; open weights and workstation-scale inference give developers control over where the model runs how it is customized and whether sensitive code or data has to leave their own environment. Meta’s bet is that the next phase of AI will not belong entirely to the cloud.
The model is huge. The important change is that more of the AI stack can now move onto hardware its users control.
