Valkyrie by AzumoSchedule a call
Open LLM trends · updated Oct 2, 2026

Which open-weight LLMs get downloaded most

A daily read of Hugging Face's top 5,000 models, focused on open LLMs: the most downloaded, what is rising, what people fine-tune, and each model family's share.

Top open LLMs

The most-downloaded open-weight LLMs

Original uploads released since 2024, ranked by downloads over the last 30 days. Re-uploads, compressed copies and fine-tunes that name another organisation's model as their base are left out. Small models rank high partly because automated test pipelines fetch them.

#ModelMakerSizeLicenceLast 30 daysDownloadsValkyrie
1Qwen3-0.6BQwen0.6BApache-2.029.4MFine-tune one like it →Fine-tune one →
2Qwen3-VL-8B-InstructQwen8BApache-2.016.1MFine-tune one like it →Fine-tune one →
3gemma-4-26B-A4B-itgoogle26B (4B active)Apache-2.013.1MFine-tune one like it →Fine-tune one →
4Qwen3-8BQwen8BApache-2.011.4MFine-tune one like it →Fine-tune one →
5gemma-4-31B-itgoogle31BApache-2.09.8MFine-tune one like it →Fine-tune one →
6Qwen2.5-7B-InstructQwen7BApache-2.09.5MFine-tune one like it →Fine-tune one →
7Qwen3.5-9BQwen9BApache-2.09.3MFine-tune one like it →Fine-tune one →
8Qwen2.5-0.5B-InstructQwen0.5BApache-2.08.7MFine-tune one like it →Fine-tune one →
9Llama-3.2-1B-Instructmeta-llama1BLlama licence7.7MFine-tune one like it →Fine-tune one →
10Qwen2.5-1.5B-InstructQwen1.5BApache-2.07.6MFine-tune one like it →Fine-tune one →
11Qwen3.5-4BQwen4BApache-2.07.5MFine-tune one like it →Fine-tune one →
12Qwen3.8-27BQwen27BApache-2.07.0MFine-tune one like it →Fine-tune one →
13gpt-oss-20bopenai20BApache-2.06.7MFine-tune one like it →Fine-tune one →
14Qwen3-4BQwen4BApache-2.06.7MFine-tune one like it →Fine-tune one →
15Llama-3.1-8B-Instructmeta-llama8BLlama licence6.2MFine-tune one like it →Fine-tune one →
16Qwen2.5-VL-7B-InstructQwen7BApache-2.06.1MFine-tune one like it →Fine-tune one →
17GLM-5.3-Flashzai-org321BMIT5.1MFine-tune one like it →Fine-tune one →
18Qwen3.5-2BQwen2BApache-2.04.9MFine-tune one like it →Fine-tune one →
19NVIDIA-Nemotron-3-Nano-4B-BF16nvidia4BCustom licence4.8M
20DeepSeek-V4-Flash-0731deepseek-ai304BMIT4.5MFine-tune one like it →Fine-tune one →
21gpt-oss-120bopenai120BApache-2.04.5MFine-tune one like it →Fine-tune one →
22gemma-4-E4B-itgoogle4B effective (8B total)Apache-2.04.4MFine-tune one like it →Fine-tune one →
23Qwen3-32BQwen32BApache-2.04.1MFine-tune one like it →Fine-tune one →
24Qwen2.5-3B-InstructQwen3BCustom licence4.1M
25Qwen3-4B-Instruct-2507Qwen4BApache-2.03.8MFine-tune one like it →Fine-tune one →
Rising this week

New models breaking out

Open LLMs released in the last 60 days with the biggest ranking jumps since Sep 25, 2026, among those with 25 or more likes or a trending score of 10 or more. Ranks are among Hugging Face's top 5,000 models.

MiMo-V2.6-Pro-RLXiaomiMiMoNew to the top 5,000, now #2,987Trending score 135Fine-tune one like it →Fine-tune one →
Qwen3.8-Flash-NextQwen#522 to #333Trending score 116
MiniCPM5-2Bopenbmb#660 to #433Trending score 28Fine-tune one like it →Fine-tune one →
GLM-5.3zai-org#379 to #289Trending score 98
needle3Cactus-Compute#3,251 to #2,480Trending score 38Fine-tune one like it →Fine-tune one →
DeepSeek-V4.1-Flashdeepseek-ai#675 to #525Trending score 170Fine-tune one like it →Fine-tune one →
MiMo-V2.6-Flash-RLXiaomiMiMoNew to the top 5,000, now #3,994Trending score 79Fine-tune one like it →Fine-tune one →
What teams build on

The base models people fine-tune and customise

Base models with the most popular fine-tuned and custom versions published in the last six months.

Custom and fine-tuned versionsCompressed copies
Qwen3.8-27B2 custom, 74 compressed
Qwen3.6-35B-A3B2 custom, 22 compressed
gemma-4-31B-it2 custom, 3 compressed
gemma-4-E4B-it2 custom, 3 compressed
Qwen3.6-27B1 custom, 31 compressed
gemma-4-12B-it1 custom, 17 compressed
Qwen3.5-4B1 custom, 3 compressed
Qwen3.5-0.8B1 custom, 2 compressed
Qwen3-4B1 custom, 1 compressed
SmolLM3-3B1 custom, 0 compressed

Popular versions only (those in Hugging Face's top 5,000), created in the last six months. Re-uploads and compressed copies count as compressed, and draft models are left out; fine-tunes with a compressed format in their name also count as compressed, so customisation is understated.

Who's winning

Share of open LLM downloads by model family

Weekly share since Mar 16, 2026. Copies and fine-tunes that name their base model on Hugging Face count toward that base's family; anything else counts under its uploader, usually in Other.

Qwen 48.8%Gemma 10.7%Llama 3.6%DeepSeek 3.6%gpt-oss 1.6%Other 31.8%

Shaded: Jun 10 to 19, when almost every model's reported 30-day count on Hugging Face dipped and then recovered.

Show the numbers
Week ofQwenGemmaLlamaDeepSeekgpt-ossOther
Mar 1648.4%3.6%8.2%3.5%2.9%33.4%
Mar 2348.3%3.5%8.4%3.4%2.7%33.6%
Mar 3048.4%3.6%8.4%3.5%2.5%33.6%
Apr 647.9%5.3%8.1%3.7%2.3%32.8%
Apr 1347.4%7.2%7.6%5.0%2.2%30.6%
Apr 2052.5%8.1%6.4%4.8%1.9%26.3%
Apr 2756.9%8.5%5.5%4.4%1.7%23.0%
May 457.7%8.8%5.4%4.4%1.7%21.9%
May 1157.7%9.2%5.3%4.2%1.7%21.9%
May 1853.9%10.1%5.6%4.3%1.8%24.3%
May 2548.3%11.5%5.9%5.2%2.0%27.1%
Jun 147.0%12.0%5.8%5.6%1.9%27.8%
Jun 846.7%12.5%5.6%5.7%1.7%27.7%
Jun 1546.5%13.6%5.4%5.9%1.6%27.0%
Jun 2246.8%14.4%5.3%5.7%1.6%26.1%
Jun 2946.3%15.2%5.4%5.6%1.6%25.9%
Jul 646.1%15.5%5.2%5.5%1.6%26.1%
Jul 1345.7%15.6%5.1%5.2%1.6%26.8%
Jul 2045.7%15.0%5.0%4.9%1.6%27.8%
Jul 2746.2%14.3%4.8%4.5%1.6%28.6%
Aug 346.4%13.8%4.7%4.4%1.6%29.1%
Aug 1047.5%13.1%4.5%4.2%1.6%29.1%
Aug 1748.4%12.6%4.1%3.9%1.6%29.4%
Aug 2449.2%12.4%3.7%3.7%1.6%29.4%
Aug 3150.1%11.3%3.4%3.4%1.6%30.2%
Sep 750.5%10.8%3.3%3.2%1.6%30.6%
Sep 1449.9%10.7%3.3%3.3%1.6%31.2%
Sep 2149.9%10.5%3.5%3.4%1.6%31.1%
Sep 2848.8%10.7%3.6%3.6%1.6%31.8%

Licences of the most-downloaded open LLMs

Apache-2.0 66.8%MIT 13.0%Llama licence 3.9%Gemma licence 2.3%Other or none 14.0%

As declared on each repo, copies included.