Best Open-Source LLMs Ranked (2026)

What is the best open-source LLM right now?

Kimi K3.1 Pro, from Moonshot AI, is the best open-weight model by Artificial Analysis' Intelligence Index, at 46 as of Nov 13, 2026. The next downloadable models are well behind: DeepSeek V4.1 Flash 1017 scores 33 and GLM-5.3 scores 32. Moonshot published K3.1 Pro's weights in a Hugging Face repository created Nov 2, 2026, a week after Artificial Analysis had listed it as "weights not yet released." It is a 2.8-trillion-parameter mixture-of-experts model that uses 104 billion parameters per token and supports a one-million-token context.

On Arena, where people vote between two anonymous answers, the top open model is the earlier Kimi K3 (Max) at #20 with a score of 1487. Kimi K3.1 Pro is #24, GLM-5.3 #28 and Xiaomi's MiMo-V2.6-Pro #31. Only seven of Arena's top 50 text models are open-weight, and all seven come from Chinese labs. That is the picture in Figure 1, Seven Seats in Fifty.

How far behind closed models are open-source LLMs?

The best open model trails the best closed model by 19 points on Artificial Analysis' index: 46 for Kimi K3.1 Pro vs. 65 for Claude Fable 5.2 (max). The gap jumped when Artificial Analysis launched version 5 of its index on Oct 27, 2026. The new version is built from nine harder evaluations split evenly across knowledge work, coding, science and general tasks, and it widened the open-weights gap from 3 points to 12 at launch, before Claude Fable 5.2 and Gemini 4 Carbon pushed the closed lead further.

Measured in time, Epoch AI found that from January to May 2026 the best open models lagged the best closed ones by an average of four months, an 8-point gap on its capability index. On Arena the distance is smaller: Kimi K3 sits 39 points below the #1 model, Gemini 4 Carbon (High).

Which open-source LLM is best for coding?

On vendor-reported coding benchmarks, DeepSeek V4 Pro scores 93.5% on LiveCodeBench and 80.6% on SWE Verified, and Kimi K3 scores 88.3 on Terminal-Bench 2.1. Treat these as best-case numbers. DeepSeek's figures are from its Max reasoning mode. In non-thinking mode the same model scores 56.8% on LiveCodeBench, 36.7 points lower (Figure 4). Kimi K3's Terminal-Bench score uses Moonshot's own Kimi Code harness. Each lab runs its own harness and settings, so cross-vendor tables are rough at best.

How much cheaper are open-source LLM APIs?

DeepSeek charges $1.20 per million output tokens for DeepSeek V4.1 Flash at peak hours, 1/33 of Claude Fable 5.2's $40 and 1/42 of GPT-6 Astra's $50. Z.ai charges $4.40 for GLM-5.3, under a quarter of Claude Opus 5.5's $20. OpenRouter's usage data shows where developers went: open-weight models handled 72% of OpenRouter tokens in the week of Sep 14, 2026, and models from Chinese labs alone handled 69%.

Are Chinese open-source models the most downloaded?

Yes. Chinese open models reached 1.15 billion cumulative Hugging Face downloads by March 2026, vs. 723 million for U.S. models and 163 million for European Union models, according to the ATOM Report. China overtook the U.S. in late July 2025. Chinese labs also hold every open seat in Arena's top 50 and the top three open spots on Artificial Analysis' index; the first U.S. open model there is Thinking Machines Lab's Inkling 2 Small, fourth at 31.

Can you run the best open-weight models yourself?

Only on serious hardware. The flagships are mixture-of-experts models that activate 3% to 5% of their parameters for each token, which keeps them fast. But all of the weights still have to sit in memory: 2.8 trillion parameters for Kimi K3.1 Pro, 1.6 trillion for DeepSeek V4 Pro and 744 billion for GLM-5.3. For a single machine, smaller dense models such as Google's Gemma 4 31B (30.7 billion parameters, Apache 2.0, 256K context) are the practical option.

Methodology & caveats

This graphic started from TechLatest's "Best Open-Weight LLMs in July 2026" roundup. Every model figure it quotes for Kimi K3, Kimi K2.7 Code, Inkling, GLM-5.2, MiniMax M3, DeepSeek V4 Pro and Gemma 4 31B matches the vendor's own model card, but several need context. GLM-5.2's 753B is Hugging Face's automatic weight count; Z.ai labels the model 744B total and 40B active. Its Terminal Bench 2.1 score of 82.7 is the best-harness result, while Z.ai's headline is 81.0. DeepSeek's scores are Max mode. DeepSeek V4 Pro (Apr 24) and Gemma 4 (Apr 2) were not July releases. Several of the article's models have since been superseded: GLM-5.2 by GLM-5.3 and Kimi K3 by K3.1 Pro.

Rankings are snapshots. The Arena leaderboard was last updated Nov 11, 2026, and both Artificial Analysis pages were read Nov 13, 2026. The 19-point gap and every share or ratio marked "our calculation" are derived from published figures. Artificial Analysis scores each model at a named reasoning setting, and its leaderboard and open-weights pages did not show a date. DeepSeek V4.1 Flash 1017 scored 32 in the Oct 27 launch article and 33 on the live page. Download counts end in March 2026, OpenRouter shares in September 2026 and Epoch's lag estimate in May 2026, so the measures cover different periods. "Open-weight" means the weights can be downloaded, often under custom licenses (Kimi's, MiniMax's) with conditions. It does not mean the training data or code is open. Prices are list prices for each lab's own API and exclude caching discounts and third-party hosts.