Free AI Leaderboard, Ranked by Real Votes

Only models you can use without paying: free online, or downloaded to your own GPU. Scores come from millions of blind human votes on Arena, updated every night.

Best free LLMs · Best free AI for every task · Free AI chats

On this page: 3 charts · Rankings · New models · FAQ

New on the free leaderboard

Kandinsky 5.0 Video Pro#39 in Video · 5 Nov 2026 Wan 2.2 A14B#44 in Video · 5 Nov 2026 Kandinsky 5.0 Video Lite#46 in Video · 5 Nov 2026 Mochi 1#50 in Video · 5 Nov 2026

Arena score

Higher is better

Blind human votes on Arena, chat

Gemma 4 31B1452 Gemma 4 26B A4B1437 Qwen3.8-27B1435 Qwen3.5-27B1409 Qwen3-Next-80B-A3B-Instruct1398 Qwen3.5-35B-A3B1394 GLM-4.6V1378 GLM-4.5-Air1373

Bars start at 1000. On Arena, 10 points is a small gap and 50 a clear one.

Speed on one RTX 4090

Higher is better

Tokens per second at 4-bit, estimated

Qwen3.5-35B-A3B188 tok/s Gemma 4 26B A4B152 tok/s Qwen3-Next-80B-A3B-Instruct60 tok/s Qwen3.8-27B43 tok/s Qwen3.5-27B43 tok/s Gemma 4 31B36 tok/s GLM-4.6V12 tok/s GLM-4.5-Air10 tok/s

Estimate calibrated on 44 real llama.cpp runs. About 10 tok/s is reading speed.

Memory at 4-bit

Lower is better

GB to load it, with 8K context

Gemma 4 26B A4B17 GB (fits 24 GB)Qwen3.8-27B18 GB (fits 24 GB)Qwen3.5-27B18 GB (fits 24 GB)Gemma 4 31B22 GB (fits 24 GB)Qwen3.5-35B-A3B23 GB (fits 24 GB)Qwen3-Next-80B-A3B-Instruct50 GB (GPU + RAM)GLM-4.6V67 GB (GPU + RAM)GLM-4.5-Air69 GB (GPU + RAM)

Green: fits fully on one 24 GB card. Red: also needs system RAM, so it runs slower.

What this shows · How to read it

What it shows. The 8 best free chat models that run on a home PC with one RTX 4090 (24 GB) and 64 GB of RAM, compared three ways: how good they are, how fast they write, and how much memory they need.

How to read it. Same models, same colors in all three charts, so you can follow one model across them. A model can be the best but slow, or small and fast: pick the balance you need, then check your own card in Run at home.

Free API right now Downloadable (open weights)

Arena ranking of 6 Nov 2026 · official Arena data, checked every 2 hours

228 of 419 models in Arena's Chat ranking are free to use. Best free: Kimi K3 (Max), #19 overall.

Filter by lab:

Full free ranking

Free rank	Model	Arena score	Size	Download	4-bit needs	Quality at 4-bit	Speed on RTX 4090	Use free online
1	Kimi K3 (Max) Kimi (Moonshot)	1488 #19 of 419	2779.9 B 104.6 B active	Kimi K3 license	1686 GB	+1.2% very small loss	does not fit	Lab chat
2	GLM-5.3 (Max) GLM (Z.ai)	1482 #26 of 419	753.3 B 50.2 B active	MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
3	MiMo-V2.6-Pro Xiaomi	1479 #29 of 419	1024.2 B 43.9 B active	MIT	622 GB	+1.2% very small loss	does not fit	—
4	GLM-5.3-Flash GLM (Z.ai)	1476 #33 of 419	321.3 B 24.7 B active	MIT	195 GB	+1.2% very small loss	does not fit	Lab chat
5	DeepSeek-V4.1-Flash-20260910 (Max) DeepSeek	1475 #37 of 419	—	MIT	—	—	—	Lab chat
6	GLM-5.2 (Max) GLM (Z.ai)	1475 #38 of 419	753.3 B 50.3 B active	MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
7	MiMo-V2.5-Pro Xiaomi	1466 #52 of 419	1023.2 B 42.9 B active	MIT	621 GB	+1.2% very small loss	does not fit	—
8	GLM-5.1 GLM (Z.ai)	1465 #55 of 419	753.9 B 50.8 B active	MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
9	DeepSeek-V4-Pro (High, 20260813) DeepSeek	1464 #58 of 419	1598.8 B 74.7 B active	MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
10	Kimi K2.6 Kimi (Moonshot)	1461 #60 of 419	1026.9 B 32.2 B active	Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
11	DeepSeek-V4-Pro DeepSeek	1458 #64 of 419	1598.8 B 74.7 B active	MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
12	GLM-5 GLM (Z.ai)	1457 #65 of 419	753.9 B 50.8 B active	MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
13	DeepSeek-V4-Pro Preview (High) DeepSeek	1455 #72 of 419	1598.8 B 74.7 B active	MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
14	Hy3 Tencent	1454 #75 of 419	298.8 B 23.9 B active	Apache 2.0	184 GB	+1.2% very small loss	does not fit	Lab chat
15	Gemma 4 31BBest on one 4090 Google	1452 #77 of 419	31.3 B	Apache 2.0	22 GB	+1.8% small loss	36 tok/s · fits	Free API · try it Lab chat
16	MiMo-V2.6-Flash Xiaomi	1450 #79 of 419	310.8 B 16.8 B active	MIT	189 GB	+1.2% very small loss	does not fit	—
17	Kimi K2.5 (Thinking) Kimi (Moonshot)	1450 #81 of 419	1026.9 B 32.2 B active	Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
18	Inkling Thinky	1443 #90 of 419	—	Apache 2.0	—	—	—	—
19	GLM-4.7 GLM (Z.ai)	1442 #95 of 419	358.3 B 38.4 B active	MIT	221 GB	+1.2% very small loss	does not fit	Lab chat
20	Qwen3.5-397B-A17B Qwen (Alibaba)	1442 #96 of 419	403.4 B 17 B active	Apache 2.0	245 GB	+1.2% very small loss	does not fit	Lab chat
21	DeepSeek-V4-Flash Preview (High) DeepSeek	1439 #99 of 419	290.9 B 19.9 B active	MIT	178 GB	+1.2% very small loss	does not fit	Lab chat
22	Gemma 4 26B A4BFastest on one 4090 Google	1437 #100 of 419	25.8 B 4 B active	Apache 2.0	17 GB	+1.9% small loss	152 tok/s · fits	Free API · try it Lab chat
23	MiniMax-M3 MiniMax	1436 #103 of 419	427 B 25.6 B active	Open weights MiniMax Community License	260 GB	+1.2% very small loss	does not fit	Lab chat
24	DeepSeek-V4-Flash DeepSeek	1436 #104 of 419	290.9 B 19.9 B active	MIT	178 GB	+1.2% very small loss	does not fit	Lab chat
25	Beam 1.0 501B A23B Reflection	1435 #106 of 419	—	Apache 2.0	—	—	—	—
26	Qwen3.8-27B Qwen (Alibaba)	1435 #108 of 419	27.8 B	Apache 2.0	18 GB	+1.9% small loss	43 tok/s · fits	Lab chat
27	MiMo-V2.5 Xiaomi	1434 #111 of 419	310.8 B 16.8 B active	MIT	189 GB	+1.2% very small loss	does not fit	—
28	Kimi K2.5 (Instant) Kimi (Moonshot)	1431 #114 of 419	1026.9 B 32.2 B active	Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
29	Kimi K2 Thinking Turbo Kimi (Moonshot)	1430 #117 of 419	—	Modified MIT	—	—	—	Lab chat
30	Mistral Medium 3.5 Mistral	1427 #119 of 419	—	Modified MIT	—	—	—	—
31	NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) NVIDIA	1426 #122 of 419	560.5 B 55 B active	OpenMDW-1.1	340 GB	+1.2% very small loss	does not fit	Free API · try it Lab chat
32	DeepSeek-V3.2-Exp (Thinking) DeepSeek	1425 #123 of 419	685.4 B 51 B active	MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
33	DeepSeek-V3.2 DeepSeek	1425 #124 of 419	685.4 B 51 B active	MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
34	GLM-4.6 GLM (Z.ai)	1424 #125 of 419	356.8 B 36.9 B active	MIT	220 GB	+1.2% very small loss	does not fit	Lab chat
35	Muse Glimmer Meta	1424 #126 of 419	—	Apache-2.0	—	—	—	—
36	DeepSeek-V3.2 (Thinking) DeepSeek	1422 #129 of 419	685.4 B 51 B active	MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
37	Qwen3-235B-A22B-Instruct (2507) Qwen (Alibaba)	1422 #130 of 419	—	Apache 2.0	—	—	—	Lab chat
38	DeepSeek-V3.2-Exp DeepSeek	1422 #131 of 419	685.4 B 51 B active	MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
39	DeepSeek-R1 (0528) DeepSeek	1421 #132 of 419	—	MIT	—	—	—	Lab chat
40	Mistral Large 4 Mistral	1420 #133 of 419	—	Modified MIT	—	—	—	—
41	Kimi K2 Preview (0905) Kimi (Moonshot)	1419 #134 of 419	—	Modified MIT	—	—	—	Lab chat
42	Kimi K2 Preview (0711) Kimi (Moonshot)	1418 #137 of 419	—	Modified MIT	—	—	—	Lab chat
43	DeepSeek-V3.1 DeepSeek	1417 #138 of 419	684.5 B 50.1 B active	MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
44	Qwen3.5-122B-A10B Qwen (Alibaba)	1416 #139 of 419	125.1 B 10 B active	Apache 2.0	77 GB	+1.2% very small loss	does not fit	Lab chat
45	DeepSeek-V3.1-Terminus (Thinking) DeepSeek	1416 #140 of 419	684.5 B 50.1 B active	MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
46	DeepSeek-V3.1 (Thinking) DeepSeek	1416 #141 of 419	684.5 B 50.1 B active	MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
47	MiniMax-M2.7 MiniMax	1415 #142 of 419	228.7 B 10.4 B active	Modified MIT	141 GB	+1.2% very small loss	does not fit	Lab chat
48	DeepSeek-V3.1-Terminus DeepSeek	1415 #143 of 419	684.5 B 50.1 B active	MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
49	Mistral Large 3 Mistral	1414 #147 of 419	—	Apache 2.0	—	—	—	—
50	Qwen3-VL-235B-A22B-Instruct Qwen (Alibaba)	1413 #148 of 419	235.7 B 22 B active	Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
51	NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4) NVIDIA	1411 #150 of 419	124.5 B 12 B active	OpenMDW-1.1	76 GB	+1.2% very small loss	does not fit	Free API · try it Lab chat
52	GLM-4.5 GLM (Z.ai)	1411 #153 of 419	358.3 B 38.4 B active	MIT	221 GB	+1.2% very small loss	does not fit	Lab chat
53	Qwen3.5-27B Qwen (Alibaba)	1409 #155 of 419	27.8 B	Apache 2.0	18 GB	+1.9% small loss	43 tok/s · fits	Lab chat
54	Hy3 preview Tencent	1409 #156 of 419	298.8 B 23.9 B active	Open weights tencent-hunyuan-community	184 GB	+1.2% very small loss	does not fit	Lab chat
55	Inkling-Small Thinky	1405 #158 of 419	—	Apache 2.0	—	—	—	—
56	Qwen3-235B-A22B (Non-Thinking) Qwen (Alibaba)	1402 #162 of 419	—	Apache 2.0	—	—	—	Lab chat
57	LongCat-Flash-Chat Meituan	1401 #165 of 419	—	MIT	—	—	—	—
58	Qwen3-235B-A22B-Thinking (2507) Qwen (Alibaba)	1400 #166 of 419	—	Apache 2.0	—	—	—	Lab chat
59	Qwen3-Next-80B-A3B-Instruct Qwen (Alibaba)	1398 #167 of 419	81.3 B 3 B active	Apache 2.0	50 GB	+1.2% very small loss	60 tok/s GPU + RAM	Lab chat
60	DeepSeek-R1 DeepSeek	1398 #168 of 419	684.5 B 50.1 B active	MIT	416 GB	+1.2% very small loss	does not fit	Lab chat

Size: parameters (and active parameters for mixture-of-experts). 4-bit needs: GPU memory at Q4_K_M with 8K of context. Quality at 4-bit: perplexity increase against the full model, from llama.cpp tests scaled by size. Speed: estimate on one RTX 4090 (24 GB) with 64 GB of DDR5, at 4-bit. Simulate your own hardware.

Top 15 free chat models

Kimi K3 (Max) 1488 GLM-5.3 (Max) 1482 MiMo-V2.6-Pro 1479 GLM-5.3-Flash 1476 DeepSeek-V4.1-Flash-20260910 (Max) 1475 GLM-5.2 (Max) 1475 MiMo-V2.5-Pro 1466 GLM-5.1 1465 DeepSeek-V4-Pro (High, 20260813) 1464 Kimi K2.6 1461 DeepSeek-V4-Pro 1458 GLM-5 1457 DeepSeek-V4-Pro Preview (High) 1455 Hy3 1454 Gemma 4 31B 1452

140014501500

Arena score, axis starts at 1400. Color = the lab that made the model, the same in every chart. Dashed line: best paid model, Gemini 4 Carbon (High) (1526).

What this shows · How to read it

What it shows. The 15 best free models for chat, by their score from blind human votes on Arena. The dashed line is the best paid model, for reference.

How to read it. Longer bar, better model. The axis does not start at zero, so small gaps look bigger than they are: on Arena, 10 points is a small difference, 50 is a clear one. Each color is a lab; use the lab buttons above to show only some.

Score against size

Small and strong Kimi K3 (Max)GLM-5.3-FlashHy3Gemma 4 31Bgpt-oss-20bGranite 4.2 3B

130014001500

10B100B1T

Each dot is an open model; up is better, left is smaller. The line joins the models that beat every smaller one. Parameters on a log scale. Color = the lab that made the model.

What this shows · How to read it

What it shows. Every open model with a known size: its Arena score against its number of parameters. Size decides how much memory you need to run it.

How to read it. Up is better, left is smaller. Models in the green corner are small and strong: the best value to run at home. The line links the models that beat everything smaller than them.

Score against speed at home

Fast and strong Gemma 4 31BGemma 4 26B A4BQwen3.5-35B-A3BNVIDIA-Nemotron-3.5-Lightning-30B-A3B (NVFP4)Granite 4.2 3B

1300135014001450

1 tok/s10 tok/s100 tok/s

Each dot is an open model; up is better, right is faster on one RTX 4090 at 4-bit. The line joins the models that beat every faster one. Speed on a log scale. Color = the lab that made the model.

What this shows · How to read it

What it shows. Open models that run on one RTX 4090 (24 GB) with 64 GB of system RAM, at 4-bit: Arena score against the estimated words per second.

How to read it. Up is better, right is faster. The green corner is fast and strong: the best value for a home PC. Around 10 tokens per second is reading speed; below 5 feels slow.

Top free models over the last six months

14501475150015251550

JunJulAugSeptOctNov

Kimi K3 (Max) 1488GLM-5.3 (Max) 1482MiMo-V2.6-Pro 1479GLM-5.3-Flash 1476DeepSeek-V4.1-Flash-20260910 (Max) 1475Best paid model

Arena score from 12 May 2026 to 6 Nov 2026.

What this shows · How to read it

What it shows. How the Arena score of the five best open models changed over six months, against the best paid model (dashed grey).

How to read it. A line going up means the model won more votes over time. Lines start when a model joined Arena. Same color as its lab; a dashed line is a second model from the same lab.

Who beats whom

Model A ↓ · Model B →	Gemini 4 Carbon (High) (paid)	Kimi K3 (Max)	GLM-5.3 (Max)	MiMo-V2.6-Pro	GLM-5.3-Flash	DeepSeek-V4.1-Flash-20260910 (Max)	GLM-5.2 (Max)	MiMo-V2.5-Pro	GLM-5.1	DeepSeek-V4-Pro (High, 20260813)	Kimi K2.6	DeepSeek-V4-Pro	GLM-5
Gemini 4 Carbon (High) paid		0.55	0.56	0.57	0.57	0.57	0.57	0.59	0.59	0.59	0.59	0.60	0.60
Kimi K3 (Max)	0.45		0.51	0.51	0.52	0.52	0.52	0.53	0.53	0.53	0.54	0.54	0.54
GLM-5.3 (Max)	0.44	0.49		0.50	0.51	0.51	0.51	0.52	0.52	0.53	0.53	0.53	0.54
MiMo-V2.6-Pro	0.43	0.49	0.50		0.50	0.51	0.51	0.52	0.52	0.52	0.53	0.53	0.53
GLM-5.3-Flash	0.43	0.48	0.49	0.50		0.50	0.50	0.51	0.52	0.52	0.52	0.53	0.53
DeepSeek-V4.1-Flash-20260910 (Max)	0.43	0.48	0.49	0.49	0.50		0.50	0.51	0.51	0.52	0.52	0.52	0.53
GLM-5.2 (Max)	0.43	0.48	0.49	0.49	0.50	0.50		0.51	0.51	0.52	0.52	0.52	0.53
MiMo-V2.5-Pro	0.41	0.47	0.48	0.48	0.49	0.49	0.49		0.50	0.50	0.51	0.51	0.51
GLM-5.1	0.41	0.47	0.48	0.48	0.48	0.49	0.49	0.50		0.50	0.51	0.51	0.51
DeepSeek-V4-Pro (High, 20260813)	0.41	0.47	0.47	0.48	0.48	0.48	0.48	0.50	0.50		0.50	0.51	0.51
Kimi K2.6	0.41	0.46	0.47	0.47	0.48	0.48	0.48	0.49	0.49	0.50		0.50	0.51
DeepSeek-V4-Pro	0.40	0.46	0.47	0.47	0.47	0.48	0.48	0.49	0.49	0.49	0.50		0.50
GLM-5	0.40	0.46	0.46	0.47	0.47	0.47	0.47	0.49	0.49	0.49	0.49	0.50

A usually loses (down to 0.40)Even (0.50)A usually wins (up to 0.60)

Each cell: how often the model on the row is expected to beat the model in the column, from their Arena scores (1 / (1 + 10^((B − A) / 400)), the Elo formula behind Arena's ratings). Arena does not publish every single battle, so these are expected rates, not counted ones.

What this shows · How to read it

What it shows. For each pair of top models, how often the one on the row is expected to beat the one in the column in a blind vote.

How to read it. Read across a row: blue cells are matches that model usually wins, red ones it usually loses. 0.50 is a coin toss; 0.60 means it wins 6 times out of 10.

Small models, no Arena votes yet

Official models from the labs' Hugging Face pages, 0.5 to 16 billion parameters. Arena has not voted on them, so they are ordered by downloads, which shows popularity, not quality. They run on 4 to 12 GB of GPU memory.

Model	Downloads/month	Size	Run at home
Qwen3-0.6B Qwen	21.0M	0.8 B	~3 GB at 4-bit · simulate
Qwen3-14B Qwen	12.9M	14.8 B	~12 GB at 4-bit · simulate
Qwen3-VL-8B-Instruct Qwen	7.1M	8.8 B	~8 GB at 4-bit · simulate
Qwen3-8B Qwen	7.0M	8.2 B	~7 GB at 4-bit · simulate
Qwen3.5-9B Qwen	6.3M	9.7 B	~8 GB at 4-bit · simulate
Qwen2.5-VL-7B-Instruct Qwen	6.0M	8.3 B	~8 GB at 4-bit · simulate
NVIDIA-Nemotron-3-Nano-4B-BF16 nvidia	1.9M	4 B	~5 GB at 4-bit · simulate
SmolVLM2-500M-Video-Instruct HuggingFaceTB	1.8M	0.5 B	~2 GB at 4-bit · simulate
MiniCPM5-2B openbmb	1.1M	2.5 B	~4 GB at 4-bit · simulate
gemma-3-4b-it google · license to accept	1.0M	4.3 B	~5 GB at 4-bit · simulate
DeepSeek-R1-Distill-Qwen-1.5B deepseek-ai	666K	1.8 B	~3 GB at 4-bit · simulate
granite-4.1-3b ibm-granite	529K	3.4 B	~4 GB at 4-bit · simulate
SmolLM3-3B HuggingFaceTB	433K	3.1 B	~4 GB at 4-bit · simulate
gemma-3-12b-it google · license to accept	365K	12.2 B	~10 GB at 4-bit · simulate
MiniCPM-V-4.7-1B openbmb	355K	1.6 B	~3 GB at 4-bit · simulate
DeepSeek-R1-0528-Qwen3-8B deepseek-ai	355K	8.2 B	~7 GB at 4-bit · simulate
Phi-4-mini-instruct microsoft	337K	3.8 B	~5 GB at 4-bit · simulate
DeepSeek-R1-Distill-Qwen-7B deepseek-ai	331K	7.6 B	~7 GB at 4-bit · simulate
Apertus-8B-Instruct-2509 swiss-ai	309K	8.1 B	~7 GB at 4-bit · simulate
MiniCPM5-1B openbmb	239K	1.1 B	~3 GB at 4-bit · simulate

New, waiting for Arena votes

Released in the last 60 days by the labs' official Hugging Face pages, or new free models on OpenRouter, that Arena has not voted on yet. We check every night; once Arena votes, a model moves into the ranking above.

Released	Model	Type	Size	4-bit needs	Use it free
30 Oct 2026	Mage-VL-Cowork microsoft	Chat, Vision	4.7 B	5 GB	Download
27 Oct 2026	Ling-3.1-flash-1019 inclusionAI	Chat	572.1 B 32.4 B active	383 GB	Download
24 Oct 2026	d1-2B LiquidAI	Chat, Vision	1.9 B	3 GB	Download
21 Oct 2026	Flint RecursiveSI	Chat	39.8 B 2.5 B active	29 GB	Free API · try itDownload
19 Oct 2026	Ling-3.0-flash-Sante inclusionAI	Chat	127.5 B 8.2 B active	87 GB	Download
19 Oct 2026	Fibo-2.0-edit-turbo briaai	Image edit	8.1 B	7 GB	Download
18 Oct 2026	Fibo-2.0-dev briaai	Image	8.1 B	7 GB	Download
17 Oct 2026	Sentiment-Analysis-7B microsoft	Chat	7.2 B	7 GB	Download
17 Oct 2026	DeepSeek-V4.1-Flash-1017 deepseek-ai	Chat, Vision	763.2 B 227.5 B active	511 GB	Download
16 Oct 2026	NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 nvidia	Chat	4 B	5 GB	Download
16 Oct 2026	ZwZ2-27B inclusionAI	Chat, Vision	27.4 B	20 GB	Download
16 Oct 2026	ZwZ2-9B inclusionAI	Chat, Vision	9.4 B	8 GB	Download
15 Oct 2026	Hy-Memory-30B-A3B tencent	Chat	30.1 B 3 B active	22 GB	Download
14 Oct 2026	UserRepGen-4B microsoft	Chat	4.5 B	5 GB	Download
14 Oct 2026	granite-switch-4.1-8b ibm-granite	Chat	8.8 B	8 GB	Download
14 Oct 2026	granite-switch-4.1-30b ibm-granite	Chat	30 B	22 GB	Download
14 Oct 2026	flux-3-image-klein-pup black-forest-labs	Image	9.4 B	8 GB	Download
13 Oct 2026	flux-3-image-klein black-forest-labs	Image edit	5.5 B	6 GB	Download
9 Oct 2026	NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 nvidia	Chat, Vision	74.9 B 9 B active	52 GB	Download
7 Oct 2026	MiniCPM-V-4.7-35B-A3B openbmb	Chat, Vision	35.2 B 3 B active	25 GB	Download
1 Oct 2026	UserLM-2.0 microsoft	Chat	8 B	7 GB	Download
16 Sept 2026	Xing4.0 29B A4B XingChen-AGI	Chat	31.2 B 4 B active	23 GB	Free API · try itDownload
10 Sept 2026	fibo-scene-analyzer briaai	Chat, Vision	36 B 4.2 B active	26 GB	Download
9 Sept 2026	Kandinsky-6.0-Lite-distill-5s-Diffusers kandinskylab	Video	3.2 B	4 GB	Download
9 Sept 2026	Kandinsky-6.0-Pro-distill-5s-Diffusers kandinskylab	Video	30.1 B	22 GB	Download
free API	Crouton Alpha stealth	Chat	—	—	Free API · try it
free API	Dots Studio: Dots3-Note Preview dots-studio	Chat	—	—	Free API · try it
free API	Cohere: North Mini Code cohere	Chat	—	—	Free API · try it
free API	NVIDIA: Nemotron 3 Nano Omni nvidia	Chat	—	—	Free API · try it

Frequently asked questions

What about brand-new models?

Arena takes a few days or weeks to collect enough votes. Meanwhile, new models from the labs’ official pages appear in “New, waiting for Arena votes”, with their size, memory needs and speed. When Arena votes, they move into the ranking automatically.

What counts as a free model?

A model with open weights that you can download and run yourself, or a model you can call through a free API today (we check OpenRouter every night). Paid-only models are left out of the rankings, but we show the best paid score as a reference line.

Where do the scores come from?

From Arena (formerly LMSYS Chatbot Arena), where people compare two anonymous answers and vote for the better one. Arena publishes its leaderboard as an open dataset under the CC BY 4.0 license; we filter it to free models. We are not affiliated with Arena.

How often is it updated?

Every night. A scheduled job downloads the latest Arena data, the free model list from OpenRouter and model sizes from Hugging Face, then rebuilds the whole site, including the weekly report on the blog.

Can I run these models on my computer?

Open models, yes, if they fit in your GPU memory. The Run at home page estimates memory and speed for your exact card, from one to ten cards, using each model's real architecture.

Why do small models have no score?

Arena votes mostly on large models. We list the most popular small models from the labs' official Hugging Face pages separately and mark them “No Arena votes”, so you can still find something that runs on a 4 to 12 GB card.

Share this page:X LinkedIn Reddit WhatsApp

Spotted a wrong number or a broken link? Report a mistake