Best open-weight LLMs: how to choose
Newest frontier open-weight model: Muse Spark Open. Largest context window: GLM-5.3. Lowest API price, if you do not self-host: MiniMax M3.1 Flash. Sized for your own hardware: Nemotron 3.5 Bolt E4B (4GB). "Open weights" means the weights are downloadable; licence terms differ, so read the licence before commercial use. We have not run these models.
Reported Updated · Markdown · All models
Pick by what limits you
Newest frontier open-weight model
Released Oct 15, 2026 under Muse Model License (limited commercial terms); 816B parameters (42B active). Source
Largest context window
1,048,576 tokens, the largest published among recent open-weight models here. Source
Lowest API price, if you do not self-host
$0.15 in / $0.6 out per 1M tokens on the vendor API. Source
Sized for your own hardware
Newest model here that we file under "runs on your own hardware" (8B parameters); OpenMDW-1.1. Source
Each pick follows one stated rule over our sourced data, so it moves when the data does. This is a way to choose, not a quality ranking: we have not run these models.
All models in scope (20)
Every language model in our data with downloadable weights, newest first. Our data is not every model on the market.
| Model | Licence | Context | Price (listed) | Released | Status |
|---|---|---|---|---|---|
| Nemotron 3.5 Bolt E4B (4GB)NVIDIA | OpenMDW-1.1source | 64K tokenssource | Not published | Oct 23, 2026source | Generally available |
| MiniMax M3.1 FlashMiniMax | MiniMax Community Licensesource | 1M tokenssource | $0.15 in / $0.6 out per 1M tokenssource | Oct 19, 2026source | Generally available |
| DeepSeek-V4.1-Flash-1017DeepSeek | MITsource | 1M tokenssource | ¥1 in / ¥4 out per 1M tokens; peak: ¥2 / ¥8source | Oct 17, 2026source | Generally available |
| Muse Spark OpenMeta | Muse Model License (limited commercial terms)source | 1M tokenssource | Not published | Oct 15, 2026source | Generally available |
| Mistral Large 4Mistral AI | Mistral Ecosystem Open-Weight License (MIT derivative with limits on model-as-a-service)source | 1M tokenssource | $1.36 in / $4.18 out per 1M tokenssource | Oct 6, 2026source | Public preview |
| DeepSeek-V4.1-FlashDeepSeek | MITsource | 1M tokenssource | ¥1 in / ¥4 out per 1M tokens; peak: ¥2 / ¥8source | Sep 10, 2026source | Generally available |
| North Small TranslateCohere | CC BY-NC 4.0source | 16K tokenssource | Not published | Sep 9, 2026source | Generally available |
| GLM-5.3Z.ai | GLM-5.3 License (MIT-style plus a MaaS condition)source | 1.05M tokenssource | $1.4 in / $4.4 out per 1M tokenssource | Aug 14, 2026source | Generally available |
| Qwen3.8-27BAlibaba | Apache-2.0source | 262K tokenssource | Not published | Aug 14, 2026source | Generally available |
| Qwen3.8-2.4T-A95BAlibaba | Qwen3.8-Max License (custom)source | 262K tokenssource | Not published | Aug 12, 2026source | Generally available |
| Kimi K3Moonshot AI | Kimi K3 License (custom)source | 1M tokenssource | $3 in / $15 out per 1M tokenssource | Jul 16, 2026source | Generally available |
| North Mini CodeCohere | Apache 2.0source | 256K tokenssource | Not published | Jun 9, 2026source | Generally available |
| Command A+Cohere | Apache 2.0source | 128K tokenssource | Not published | May 20, 2026source | Generally available |
| Mistral Medium 3.5Mistral AI | Modified MITsource | 256K tokenssource | $1.5 in / $7.5 out per 1M tokenssource | Apr 28, 2026source | Generally available |
| Qwen3.6-27BAlibaba | Apache-2.0source | 262K tokenssource | Not published | Apr 22, 2026source | Generally available |
| GLM-5.1Z.ai | MITsource | 200K tokenssource | $1.4 in / $4.4 out per 1M tokenssource | Apr 7, 2026source | Generally available |
| Gemma 4Google | Apache 2.0source | 256K tokenssource | Not published | Mar 31, 2026source | Generally available |
| Mistral Small 4Mistral AI | Apache 2.0source | 256K tokenssource | $0.15 in / $0.6 out per 1M tokenssource | Mar 16, 2026source | Generally available |
| GLM-5Z.ai | MITsource | 200K tokenssource | $1 in / $3.2 out per 1M tokenssource | Feb 12, 2026source | Generally available |
| Ministral 3 14BMistral AI | Apache 2.0source | 256K tokenssource | $0.2 in / $0.2 out per 1M tokenssource | Dec 2, 2025source | Generally available |
FAQ
What is the best open-weight LLM?
It depends on your hardware and use. Newest frontier open-weight model: Muse Spark Open. Largest context window: GLM-5.3. Lowest API price, if you do not self-host: MiniMax M3.1 Flash. Sized for your own hardware: Nemotron 3.5 Bolt E4B (4GB).
Are open-weight models open source?
Not always. Some use permissive licences such as Apache 2.0 or MIT; others add use restrictions. Each model's licence is in the table, linked to the vendor's page.
Which open-weight models can run on my own machine?
Models we file under "runs on your own hardware": Nemotron 3.5 Bolt E4B (4GB), Qwen3.8-27B, North Mini Code, Qwen3.6-27B, Gemma 4. Check the parameter count and the vendor's hardware notes.
Did you run these models?
No. This page is Reported from the vendors' model cards, licences and pricing pages.
Related
The Weekly Verdict · free
5 new-model use cases. Every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.