On 15 October 2026, Meta released Muse Spark Open, an open-weight model it describes as "the most capable US open-weight model to date". It sits alongside Meta's closed Muse Wonder models and is aimed at coding, agentic work, long multi-step tasks and complex software engineering. The weights are free to download, and the model is already available on the major clouds and a long list of hosted APIs.
Where it shines, and where it does not
Meta's own results focus on tool use, the skill an AI agent needs to call APIs, query databases and chain steps together. There, Spark Open leads. Independent tests and coding benchmarks tell a more balanced story: the strongest Chinese open models still lead on pure coding and on overall intelligence.
Higher is better. Independent results reported in mid-October 2026.
Higher is better.
The practical reading: if you are building an agent that calls many tools, such as a data assistant that queries a warehouse, checks a dashboard and writes a summary, Spark Open is a strong open choice. For heavy coding work, Kimi K3 and GLM-5.3 currently do better, and the closed frontier models are better still. See the best AI models in October 2026 for the full picture.
Where to run it
Hosted API
Easiest- Providers
- Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI
- Pay per token, nothing to install
- Try it in minutes
Best for testing and most production use.
Your cloud
Managed- Platforms
- Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry
- Stays inside your cloud account and its controls
- You pay for dedicated GPUs
For teams with strict data rules.
Your own hardware
Advanced- Engines
- vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, Unsloth
- Full control, fine-tune with PyTorch
- Needs about 1 TB of memory for the full model
For labs and large platform teams.
Meta also contributed optimised GPU kernels to SGLang, reporting kernel-level speed-ups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, and a 3% end-to-end gain on AMD's MI350.
The licence: read before you build
"Open weights" does not mean "open source". Muse Spark Open uses the Muse Model License, which Meta describes as having limited commercial terms that let many developers use it for free. The details that matter:
- Free commercial use
- You can use and modify the model commercially.
- Large model-serving businesses
- If you sell access to models and made more than $20 million in gross revenue over 12 months, you need a custom agreement with Meta.
- Attribution
- Products must display "Built with Muse Spark Open".
- Download checks
- Hugging Face downloads ask for your legal name, date of birth and company details.
Safety
Meta says the model received additional safety training after pre-training and was evaluated under its Meta Superintelligence Scaling Framework, meeting its release thresholds. A full evaluation report is promised soon. As with any open model, once the weights are public, safety depends partly on how you deploy it: add your own guardrails, logging and human review for anything that acts on real systems.
Should data teams use it?
- 1
Start with a hosted API
Run your real prompts, for example SQL generation over your schema or tool-calling tests, before thinking about hardware.
- 2
Compare on your tasks
Test it against one closed model and one other open model. Public rankings rarely match a specific job.
- 3
Check the licence
Especially the attribution rule and the $20 million threshold if you build a product.
- 4
Plan for agents
Its tool-calling strength fits assistants that query data and call internal APIs.
Check your understanding
1/3How many parameters are active for each token?
It is a mixture-of-experts model: 816B in total, 42B active.
2/3Where does Muse Spark Open lead, according to Meta?
Kimi K3 and others lead on coding and overall intelligence.
3/3What must products built on it display?
Attribution is part of the Muse Model License.
You got 0 out of 3.
Also new this month: Mistral Large 4, Europe's 1-trillion-parameter open model, and Claude Haiku 5.5, the cheapest capable closed model.
Sources: Meta's announcement "Introducing Muse Spark Open" (15 October 2026), TechRepublic's report (16 October 2026) and the Artificial Analysis leaderboard. Benchmarks are vendor-reported unless stated.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 17 October 2026. About me