From Mexico, I’ve been waiting for a strong open model that’s actually pitched at the work I do every day: writing code with agents. On October 15, Meta released the weights for Muse Spark Open, and the announcement is direct about the target. Meta says it’s “built for coding and agentic work” and that it “excels at long-horizon tasks, complex software engineering, and multi-step reasoning.”
What Meta actually shipped
Muse Spark Open is the open-weight release of Meta’s Muse Spark. Per Meta, it’s a natively multimodal mixture-of-experts model with 816B total parameters, 42B of them active, and a 1M-token context window. The weights are on Hugging Face, and Meta lists support in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth. If you don’t want to host it, Meta names API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
The coding demo is the part that got my attention. Meta says it tested agentic workflows running entirely on-device with Muse Code powered by the model, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client: bug fixes, new features and a full build. That’s a vendor demo, so I’d treat it as a claim to test, not a result.
“Open” still means serious hardware
Here’s the honest part. Meta’s headline is that it runs on a desktop, specifically two Apple M5 Ultra Mac Studios using MLX. The Hugging Face model card is blunter: running it yourself needs a GPU setup with at least 700GB of VRAM for the NVFP4 version and 2TB for BF16. So for most of us, “open” means you can choose who hosts it, inspect it and fine-tune it, not that it fits on your laptop.
- License. It ships under the Muse Model License, which Meta describes as having limited commercial terms that allow many developers to use it for free. Read it before you build a product on it.
- Guardrails. The model card recommends deploying it as part of a broader system with extra guardrails for your use case.
- Settings. The card recommends high or xhigh reasoning effort for coding and agentic tasks.
The easy way to try it
The same day, Vercel added Muse Spark Open to its AI Gateway. Vercel says you can use it as meta/muse-spark-open in coding agents including Claude Code, Codex and Cursor after running its coding-agents setup command, and that it’s offered with ZDR and no training. For me that’s the realistic path: point an agent I already use at the new model for a week and compare it on my own tasks.
My take
I care less about the “most capable” label and more about the option it creates. A strong open-weight coding model means a team can run its agent on infrastructure it controls, tune it on its own code, and not be stuck if one hosted provider changes prices or limits. Most teams won’t buy 700GB of VRAM this week. But knowing you could move your agent to weights you control is leverage, and I’d test it on a real backlog before believing any benchmark.
Featured image: “Anvil, 5 kg in a jewellery workshop,” photo by W.carter, Wikimedia Commons, CC BY-SA 4.0. Cropped and resized to 1400×900 and lightly sharpened.
