AI
aixjo
Buyer's guideUpdated 5 October 2026

Best Open-Weight LLMs in 2026

Short answer

The strongest open-weight families in October 2026 are Qwen3.8 (Alibaba, released August 2026 with a 2.4T flagship), Kimi K3 (Moonshot AI, 2.8T parameters, 1M-token context), DeepSeek (MIT-licensed repositories), Gemma (Google DeepMind, now at Gemma 4) and gpt-oss (OpenAI, Apache 2.0).

The licence is the biggest practical difference: gpt-oss and DeepSeek are the most permissive; Kimi K3 has its own licence.

Licence and size (from vendor repositories)

Evidence: Verified
Checked 5 October 2026. Verified = read on the provider's own page on the date shown. Reported = published by a third party and linked; not yet confirmed by the provider. Claimed by provider = the vendor's own claim, not independently tested. Editorial assessment = our judgement, with reasons given.
ModelMakerSizesContextLicence
Qwen3.8Alibaba (Qwen team)2.4T-A95B (MoE), 27B262,144 tokens in deployment examplesPer model, published with weights
Kimi K3Moonshot AI2.8T total, 104B active1,048,576 tokensKimi K3 License
DeepSeek V3 / V3.2-ExpDeepSeekSee model cardsSee model cardsMIT (repositories)
Gemma 4 / 3nGoogle DeepMindSmall to mid-sizedSee model cardsModel-card terms; library Apache-2.0
gpt-ossOpenAI120b (5.1B active), 20b (3.6B active)See model cardApache 2.0
Llama 2 (legacy)MetaSuperseded—Meta licence + acceptable use policy

Benchmarks are vendor claims

Evidence: Editorial assessment

Every vendor publishes benchmark tables, usually chosen to flatter its own model. We have not run our own evaluations, so we do not rank these models on quality. Test them on your own tasks; several are available through hosted APIs so you can compare before committing to hardware.

Reported but not confirmed

Evidence: Reported

Release logs reported DeepSeek V4-Flash (31 July 2026) and V4.1 Flash under MIT (10 September), GLM-5.3 under MIT (August) and a Gemma 4 update (16 July). We could not find these on the vendors' repositories, so they are not in the table above.

How to choose

  • Licence first: Apache 2.0 or MIT if you are building a commercial product and want the fewest conditions.
  • Hardware second: the 2T+ models need data-centre hardware or a host; Qwen3.8-27B, Gemma and gpt-oss-20b are easier to run.
  • Context third: Kimi K3 lists a 1M-token window; test long inputs yourself.
  • Data rules: self-hosting keeps data in your environment; hosted APIs may not.

Common questions

What is the best open-source LLM in 2026?

There is no single best. Qwen3.8 and Kimi K3 are the largest open-weight releases; gpt-oss has the simplest licence (Apache 2.0); Gemma is the easiest to run on small hardware.

Is DeepSeek open source?

DeepSeek-V3 and V3.2-Exp are published on GitHub under the MIT licence. Check each model card for the terms that apply to the weights.

Can I use these models commercially?

Often yes, but terms differ. Apache 2.0 and MIT are permissive; Kimi K3 and Meta's Llama licences add their own conditions. Read the licence shipped with the weights.

Sources