benchmarks.site

Find by task

Open models you can run yourself

Open-weights models are published for anyone to download and run โ€” on your own server, laptop or a cheap hosting provider โ€” with no lock-in. Here is how the best of them do across public leaderboards and our own tests.

How to read this

Position score (0โ€“100): 100 = first place on every leaderboard where the model appears. Open models are compared against all models, closed ones included, so a score of 60 means "upper half of the whole field".

Which models count as open

Families that publish their weights (e.g. Qwen, DeepSeek, Llama, Gemma, Mistral, Kimi, GLM, gpt-oss, Whisper). Detected by name, so check each model's license before commercial use.

Best open models overall

Open models in at least 2 rankings (public leaderboards or our own tests)
1Qwen3.6-27B9322
2Qwen3-ASR-1.7B9333
3Qwen3-Embedding-4B8621
4gpt-oss-120b7921
5Qwen3-Embedding-8B7922
6whisper-large-v3-turbo7321
7Qwen3.5-397B-A17B7353
8whisper-large-v36643
9gemma-4-31B-it6530
10Qwen3-Embedding-0.6B6421
11Kimi K36193
12DeepSeek-V4-Pro5963
13DeepSeek V4.1 Flash58133
14DeepSeek-V4-Flash5863
15Kimi K2.65763
16Mistral-Small-3.2-24B-Instruct5721
17GLM 5.3 Flash5772
18Qwen3-VL 235B-A22B5751
19OpenAI gpt-realtime-whisper5421
20Qwen3.5-35B-A3B4940

Best open model per task

Read text from photos & documents

  1. Kimi K31001 board
  2. Kimi K2.6922 boards
  3. gemma-4-31B-it-turbo901 board
  4. Qwen3.6-27B901 board
  5. Qwen3.5-0.8B891 board

Turn speech into text (and back)

  1. Qwen3-ASR-1.7B, raw audio1001 board
  2. Nemotron-3-Diarization1001 board
  3. Qwen3-ASR-1.7B933 boards
  4. Musci-ASR-qwen-1019-hf841 board
  5. whisper-large-v3-turbo732 boards

Chat, answer & summarize

  1. Phi-4881 board
  2. Deepseek V4p1 Flash841 board
  3. Llama-3.3-70B-Instruct-Turbo841 board
  4. gpt-oss-20b781 board
  5. gemma-3-12b-it721 board

Write and fix code

  1. Z.ai GLM-5.2821 board
  2. Kimi K3.1 Pro622 boards
  3. Kimi K3582 boards
  4. DeepSeek V4.1 Flash561 board
  5. Minimax M3p1 Flash501 board

Agents that use tools & computers

  1. GLM 5.3 Flash1001 board
  2. Kimi K3.1 Pro961 board
  3. Kimi K3932 boards
  4. KV-Ground-8B + Qwen3.5-27B Consistency Router841 board
  5. DeepSeek V4.1 Flash741 board

Math & science

  1. DeepSeek V4.1 Flash381 board
  2. Kimi K3.1 Pro312 boards

Search & "find similar"

  1. Qwen3-Embedding-0.6B1001 board
  2. Qwen3-Embedding-4B831 board
  3. Qwen3-Embedding-8B671 board
  4. embeddinggemma-300m501 board
  5. multilingual-e5-large-instruct331 board

Polish & other languages

  1. Kimi K31001 board
  2. Qwen3-ASR-1.7B, raw audio1001 board
  3. Qwen3.6-27B932 boards
  4. Qwen3-ASR-1.7B933 boards
  5. Qwen3.5-122B-A10B921 board

Speed, price & providers

  1. DeepSeek V4.1 Flash1001 board
  2. GLM 5.3 Flash851 board
  3. step-5-preview-20261008701 board
  4. MiMo V2.6 Flash651 board
  5. GLM 5.3601 board

Safety & security

  1. XekRung-1.5-27B961 board
  2. gpt-oss-120b921 board
  3. GLM 5.3841 board
  4. Feyospace-v1.1801 board
  5. DeepSeek-V4-Pro761 board

Overall rankings

  1. Kimi K2.5641 board
  2. GLM-5561 board
  3. Kimi K3.1 Pro523 boards
  4. GLM 5.3401 board
  5. qwen-3-5-397b401 board

Open models in our own tests

We measured these ourselves (mostly through an API host). Click a name to see every test it appears in.