The best open-weight AI models you can run yourself (October 2026)
The strongest AI models with downloadable weights right now, from Meta's Muse Spark Open to DeepSeek and Mellum, with sizes, licences and what hardware they need.
We may earn a commission if you sign up through links on this page. It costs you nothing and never decides what we recommend. Learn more
Open-weight models let you download an AI model and run it on your own hardware. That matters when your data can’t leave the building, when you want predictable costs, or when you need to customize the model. October 2026 was a big month for them.
Quick answer: Muse Spark Open for the most capable model you can run yourself, DeepSeek V4.1 Flash for strong results on a budget, Mellum2.1 for fast coding help, and Muse Glimmer if you want something that fits on a laptop.
The best open-weight models right now
| Model | Maker | Released | Size | Licence | Runs on |
|---|---|---|---|---|---|
| Muse Spark Open | Meta | Oct 15, 2026 | 816B total, 42B active | Meta’s licence (check terms) | Two Mac Studios (M5 Ultra), or servers |
| DeepSeek V4.1 Flash | DeepSeek | Oct 17, 2026 | Large mixture-of-experts | See the Hugging Face page | Servers |
| Mellum2.1 | JetBrains | Oct 8, 2026 | 12B total, 2.5B active | Apache 2.0 | A single GPU |
| Muse Glimmer | Meta | 2026 | Small, agent-focused | Apache 2.0 | A laptop (17 GB version) |
| Kimi K3.1 Pro | Moonshot AI | Oct 23, 2026 | 3T class | Weights promised by Nov 4 | Servers |
From each maker’s own announcement, checked October 28, 2026. “Active” parameters are the part of the model used for each word, which is what drives speed.
Muse Spark Open: the most capable
Meta released Muse Spark Open on October 15, 2026, and calls it the most capable US open-weight model to date. It’s built for coding and long agent tasks, handles images as well as text, and reads up to 1 million tokens. Meta worked with Apple so it runs on two Mac Studios with M5 Ultra chips; it’s also hosted by Groq, Together AI, Fireworks AI, OpenRouter and others if you want to try before you download.
DeepSeek V4.1 Flash: best value
DeepSeek V4.1 Flash’s October 17 update is faster, gives shorter answers and scores higher on DeepSeek’s benchmarks. The weights are on Hugging Face and ModelScope, and if you’d rather not host it, DeepSeek’s API is one of the cheapest around. See Kimi vs DeepSeek for a direct comparison.
Mellum2.1: small and fast for coding
JetBrains’ Mellum2.1 has 12 billion parameters but only 2.5 billion active at a time, so it’s quick. It’s built for code completion and coding agents and released under the permissive Apache 2.0 licence.
Muse Glimmer: fits on a laptop
Meta’s Muse Glimmer is a smaller open model for personal agents, released under Apache 2.0. Meta says a 17 GB quantized version runs fast enough for real-time use on a MacBook with an M4 Max or M5 Max, or a PC with an RTX 5090.
Coming soon
- Kimi K3.1 Pro: Moonshot says the full weights will be out by November 4, 2026.
- MiniMax M3.1 Flash: MiniMax says the weights will be released later; the model is already available through its API.
Before you download
- Check the licence. Apache 2.0 lets you do almost anything; custom licences may restrict commercial use.
- Check the hardware. Big models need far more memory than a normal PC. Start with a hosted version to test quality.
- Check the quality on your task. Open models have closed much of the gap, but test on your own examples before you switch.
Track every new release, open or closed, on our new models page.
Tools in this post
DeepSeek
by DeepSeek
Free chat assistant from Chinese lab DeepSeek, strong at reasoning and coding.
Free app and web chat; paid API for developers
Privacy & risks
- Trains on your data
- Yes
- Minimum age
- Intended for adults; under 18 needs guardian consent
- Its privacy policy says user data is used to train its models.
- Data is processed by a Chinese company. Check your employer's or school's rules before using it with work or personal data.
Kimi
by Moonshot AI
Assistant for long documents, research and coding, powered by the open-weight Kimi K3.1 models.
Questions people ask
What is an open-weight AI model?
A model whose trained weights you can download and run on your own computers or servers, instead of only using it through the maker's app or API. Licences differ, so check what each one allows.
What is the best open-weight model right now?
Meta calls Muse Spark Open, released October 15, 2026, the most capable US open-weight model to date. DeepSeek V4.1 Flash is the strongest low-cost option with weights already public.
Can I run an open-weight model on my laptop?
Small ones, yes. Meta says a 17 GB version of Muse Glimmer runs on a MacBook with an M4 Max or M5 Max chip, or a PC with an RTX 5090. The biggest models need much more memory.
