Agentic WeeklyGet the weekly

Best open-source LLM right now (October 2026)

Among the 17 open-weights language models in our data, DeepSeek-V4.1-Flash-1017 ranks highest on LMArena WebDev (#14 on the LMArena WebDev snapshot of Oct 22, 2026) and is MIT-licensed. Kimi K3 and GLM-5.3 follow but add commercial conditions. For one machine, our pick is Qwen3.8-27B (Apache-2.0, #32). "Open source" and "open weights" differ: we say open weights unless the licence is a standard permissive one.

By Agentic Weekly, operated independently from Texas, US · Last updated (America/Chicago) · How we work

Free weekly brief: email newsletter · RSS

Evidence: Reported. Licences are read from each model's vendor or Hugging Face page and ranks are the LMArena WebDev snapshot of Oct 22, 2026 (coding preference only) read from the sources listed below on 2026-10-27 (America/Chicago). We did not run these products or models for this page.

Cite this: Agentic Weekly, "Best open-source LLM right now (October 2026)", https://theagenticweekly.com/what-is/best-open-source-llm, read Oct 27, 2026.

Method: Licences are read from each model's vendor or Hugging Face page; ranks are the LMArena WebDev snapshot of Oct 22, 2026, a coding-preference score only. Each fact is copied from the vendor page linked in its row into our model records with the date we read it. Reported means read from that page, not run by us; Tested would mean we ran it and logged a dated run, and no model on this page is Tested. 64 language models are in our data: 41 have a sourced USD list price, 2 are priced in another currency and 21 have no verified price. Only models with a sourced price are listed; we do not estimate or pad.

Data: models.csv · models.json · answers.json

Open-weights LLMs in our data

Sorted by LMArena WebDev rank where the model appears; models not on that board are listed after.

Open-weights language models with licence, price and WebDev rank
ModelVendorLicenceStandard permissive?WebDev rankAPI price (per 1M, in / out)Source
DeepSeek-V4.1-Flash-1017DeepSeekMITYes#14CNY 1 / 4licence source
Kimi K3Moonshot AIKimi K3 License (custom)No: read the licence#17$3 / $15licence source
MiniMax M3.1 FlashMiniMaxMiniMax Community LicenseNo: read the licence#34$0.15 / $0.60licence source
Command A+CohereApache 2.0Yesnot on the boardnone recordedlicence source
DeepSeek-V4.1-FlashDeepSeekMITYesnot on the boardCNY 1 / 4licence source
Gemma 4GoogleApache 2.0Yesnot on the boardnone recordedlicence source
GLM-5.3Z.aiGLM-5.3 License (MIT-style plus a MaaS condition)No: read the licencenot on the board$1.40 / $4.4licence source
Ministral 3 14BMistral AIApache 2.0Yesnot on the board$0.20 / $0.20licence source
Mistral Medium 3.5Mistral AIModified MITNo: read the licencenot on the board$1.50 / $7.50licence source
Mistral Small 4Mistral AIApache 2.0Yesnot on the board$0.15 / $0.60licence source
Muse Spark OpenMetaMuse Model License (limited commercial terms)No: read the licencenot on the boardnone recordedlicence source
Nemotron 3.5 Bolt E4B (4GB)NVIDIAOpenMDW-1.1No: read the licencenot on the boardnone recordedlicence source
North Mini CodeCohereApache 2.0Yesnot on the boardnone recordedlicence source
North Small TranslateCohereCC BY-NC 4.0No: read the licencenot on the boardnone recordedlicence source
Qwen3.6-27BAlibabaApache-2.0Yesnot on the boardnone recordedlicence source
Qwen3.8-2.4T-A95BAlibabaQwen3.8-Max License (custom)No: read the licencenot on the boardnone recordedlicence source
Qwen3.8-27BAlibabaApache-2.0Yesnot on the boardnone recordedlicence source

How to choose

  • Rank is coding-task preference only. We have no verified open-model board for writing or research.
  • Check the licence file before shipping: Kimi K3 and GLM-5.3 add commercial conditions in our records.
  • Running a model yourself costs compute, not tokens. We have not run these models.

FAQ

What is the best open source LLM right now?

DeepSeek-V4.1-Flash-1017 (MIT) is the highest-ranked open-weights model in our data, #14 on LMArena WebDev (Oct 22, 2026). Next: Kimi K3 (#17) and GLM-5.3. That is coding preference only, not our test.

What is the best open model to run on one machine?

Qwen3.8-27B (Apache-2.0) is our pick, then Gemma 4. We have not run either; hardware needs depend on quantization.

Is open source the same as open weights?

No. Open weights means the trained weights can be downloaded; the licence decides what you may do with them. Only MIT and Apache-2.0 rows in the table are standard permissive licences; the rest add conditions.

Sources

Page history

  • : First published. Facts read from the sources listed below on this date.

Markdown version for agents: /what-is/best-open-source-llm.md