The best free AI models, 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.

New on the free leaderboardDeepSeek-V4.1-Flash-20261017 (Max)#12 in Coding · 18 Oct 2026 DeepSeek-V4.1-Flash-20260910 (Max)#25 in Coding · 18 Oct 2026 Kimi K3 (Max)#15 in Coding · 16 Oct 2026 MiMo-V2.6-Flash#18 in Coding · 16 Oct 2026

Free API right now Downloadable (open weights)

226 of 416 models in Arena's Chat ranking are free to use. Votes as of 15 Oct 2026. Best free: mimo-v2.6-pro, #11 overall.

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	mimo-v2.6-pro xiaomi	1491 #11 of 416	1024.2 B 44.9 B active	Yes MIT	622 GB	+1.2% very small loss	does not fit	—
2	Kimi K3 (Max) moonshot	1475 #25 of 416	2779.9 B 105.8 B active	Yes Kimi K3 license	1686 GB	+1.2% very small loss	does not fit	Lab chat
3	glm-5.3-max zai	1471 #27 of 416	753.3 B 51.2 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
4	glm-5.3-flash zai	1471 #29 of 416	321.3 B 25.4 B active	Yes MIT	195 GB	+1.2% very small loss	does not fit	Lab chat
5	glm-5.2-max zai	1470 #31 of 416	753.3 B 51.2 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
6	mimo-v2.5-pro xiaomi	1465 #37 of 416	1023.2 B 43.8 B active	Yes MIT	621 GB	+1.2% very small loss	does not fit	—
7	deepseek-v4.1-flash-max deepseek	1463 #38 of 416	763.2 B 228.1 B active	Yes MIT	464 GB	+1.2% very small loss	does not fit	Lab chat
8	glm-5.1 zai	1462 #39 of 416	753.9 B 51.7 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
9	kimi-k2.6 moonshot	1455 #44 of 416	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
10	mimo-v2.6-flash xiaomi	1455 #45 of 416	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
11	deepseek-v4-pro deepseek	1451 #49 of 416	1598.8 B 75.6 B active	Yes MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
12	glm-5 zai	1446 #60 of 416	753.9 B 51.8 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
13	deepseek-v4-pro-high-20260813 deepseek	1446 #61 of 416	—	Yes MIT	—	—	—	Lab chat
14	kimi-k2.5-thinking moonshot	1445 #62 of 416	—	Yes Modified MIT	—	—	—	Lab chat
15	deepseek-v4-pro-high-preview deepseek	1445 #63 of 416	—	Yes MIT	—	—	—	Lab chat
16	nvidia-nemotron-3-ultra-550b-a55b-nvfp4 nvidia	1445 #64 of 416	560.5 B 55 B active	Yes OpenMDW-1.1	340 GB	+1.2% very small loss	does not fit	Lab chat
17	gemma-4-31b google	1443 #67 of 416	31.3 B	Yes Apache 2.0	22 GB	+1.8% small loss	33 tok/s	Free API · try it Lab chat
18	inkling thinky	1441 #74 of 416	—	Yes Apache 2.0	—	—	—	Free API · try it
19	qwen3.8-27b alibaba	1441 #75 of 416	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
20	glm-4.6 zai	1440 #76 of 416	356.8 B 37.6 B active	Yes MIT	220 GB	+1.2% very small loss	does not fit	Lab chat
21	hy3 tencent	1438 #80 of 416	298.8 B 24.4 B active	Yes Apache 2.0	184 GB	+1.2% very small loss	does not fit	Lab chat
22	qwen3.5-397b-a17b alibaba	1438 #81 of 416	403.4 B 17 B active	Yes Apache 2.0	245 GB	+1.2% very small loss	does not fit	Lab chat
23	glm-4.7 zai	1435 #87 of 416	358.3 B 39.1 B active	Yes MIT	221 GB	+1.2% very small loss	does not fit	Lab chat
24	gemma-4-26b-a4b google	1434 #90 of 416	25.8 B 4 B active	Yes Apache 2.0	17 GB	+1.9% small loss	238 tok/s	Free API · try it Lab chat
25	minimax-m3 minimax	1433 #91 of 416	427 B 26.8 B active	Yes MiniMax Community License	260 GB	+1.2% very small loss	does not fit	Lab chat
26	deepseek-v4-flash deepseek	1432 #92 of 416	290.9 B 20.4 B active	Yes MIT	178 GB	+1.2% very small loss	does not fit	Lab chat
27	glm-4.5 zai	1430 #95 of 416	358.3 B 39.1 B active	Yes MIT	221 GB	+1.2% very small loss	does not fit	Lab chat
28	mistral-large-3 mistral	1428 #98 of 416	—	Yes Apache 2.0	—	—	—	—
29	mimo-v2.5 xiaomi	1428 #100 of 416	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
30	deepseek-r1-0528 deepseek	1428 #101 of 416	684.5 B 51 B active	Yes MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
31	deepseek-v3.2-exp-thinking deepseek	1425 #104 of 416	685.4 B 51.9 B active	Yes MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
32	deepseek-v3.2 deepseek	1424 #105 of 416	685.4 B 51.9 B active	Yes MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
33	deepseek-v4-flash-high-preview deepseek	1424 #106 of 416	—	Yes MIT	—	—	—	Lab chat
34	longcat-flash-chat meituan	1423 #109 of 416	561.9 B 33.4 B active	Yes MIT	341 GB	+1.2% very small loss	does not fit	—
35	deepseek-v3.2-exp deepseek	1422 #111 of 416	685.4 B 51.9 B active	Yes MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
36	mistral-medium-3.5 mistral	1421 #112 of 416	—	Yes Modified MIT	—	—	—	—
37	kimi-k2.5-instant moonshot	1420 #113 of 416	—	Yes Modified MIT	—	—	—	Lab chat
38	qwen3-vl-235b-a22b-instruct alibaba	1420 #114 of 416	235.7 B 22 B active	Yes Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
39	deepseek-v3.2-thinking deepseek	1420 #115 of 416	685.4 B 51.9 B active	Yes MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
40	deepseek-v3.1 deepseek	1420 #116 of 416	684.5 B 51.1 B active	Yes MIT	416 GB	+1.2% very small loss	does not fit	Lab chat
41	qwen3-235b-a22b-instruct-2507 alibaba	1419 #119 of 416	235.1 B 22 B active	Yes Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
42	deepseek-v3.1-terminus-thinking deepseek	1419 #120 of 416	—	Yes MIT	—	—	—	Lab chat
43	qwen3.5-122b-a10b alibaba	1417 #123 of 416	125.1 B 10 B active	Yes Apache 2.0	77 GB	+1.2% very small loss	does not fit	Lab chat
44	deepseek-v3.1-terminus deepseek	1417 #125 of 416	—	Yes MIT	—	—	—	Lab chat
45	qwen3-next-80b-a3b-instruct alibaba	1417 #126 of 416	—	Yes Apache 2.0	—	—	—	Lab chat
46	deepseek-v3.1-thinking deepseek	1417 #128 of 416	—	Yes MIT	—	—	—	Lab chat
47	qwen3-235b-a22b-thinking-2507 alibaba	1415 #131 of 416	—	Yes Apache 2.0	—	—	—	Lab chat
48	kimi-k2-thinking-turbo moonshot	1414 #133 of 416	—	Yes Modified MIT	—	—	—	Lab chat
49	Inkling Small thinky	1413 #135 of 416	—	Yes Apache 2.0	—	—	—	Free API · try it
50	mimo-v2-flash (non-thinking) xiaomi	1411 #139 of 416	—	Yes MIT	—	—	—	—
51	nvidia-nemotron-3.5-super-vl-120b-a12b-nvfp4 nvidia	1409 #141 of 416	—	Yes OpenMDW-1.1	—	—	—	Lab chat
52	qwen3.5-27b alibaba	1409 #142 of 416	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
53	minimax-m2.7 minimax	1405 #148 of 416	—	Yes Modified MIT	—	—	—	Lab chat
54	step-3.5-flash stepfun	1403 #150 of 416	—	Yes Apache 2.0	—	—	—	—
55	hunyuan-hy3-preview tencent	1402 #151 of 416	—	Yes tencent-hunyuan-community	—	—	—	Lab chat
56	qwen3-vl-235b-a22b-thinking alibaba	1401 #152 of 416	235.7 B 22 B active	Yes Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
57	qwen3.5-35b-a3b alibaba	1395 #161 of 416	—	Yes Apache 2.0	—	—	—	Lab chat
58	mimo-v2-flash (thinking) xiaomi	1395 #162 of 416	—	Yes MIT	—	—	—	—
59	qwen3-235b-a22b-no-thinking alibaba	1393 #164 of 416	—	Yes Apache 2.0	—	—	—	Lab chat
60	minimax-m2.1-preview minimax	1392 #165 of 416	—	Yes MIT	—	—	—	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

mimo-v2.6-pro 1491 Kimi K3 (Max) 1475 glm-5.3-max 1471 glm-5.3-flash 1471 glm-5.2-max 1470 mimo-v2.5-pro 1465 deepseek-v4.1-flash-max 1463 glm-5.1 1462 kimi-k2.6 1455 mimo-v2.6-flash 1455 deepseek-v4-pro 1451 glm-5 1446 deepseek-v4-pro-high-20260813 1446 kimi-k2.5-thinking 1445 deepseek-v4-pro-high-preview 1445

1350140014501500

Arena score, axis starts at 1350. Dashed line: best paid model, Gemini 4 Argon (High) (1532).

Score against size

Small and strong mimo-v2.6-proglm-5.3-flashgemma-4-31b

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.

Top free models over the last six months

14501475150015251550

MayJunJulAugSeptOct

mimo-v2.6-pro 1491Kimi K3 (Max) 1475glm-5.3-max 1471glm-5.3-flash 1471glm-5.2-max 1470Best paid model

Arena score from 17 Apr 2026 to 15 Oct 2026.

Who beats whom

Model A ↓ · Model B →	Gemini 4 Argon (High) (paid)	mimo-v2.6-pro	Kimi K3 (Max)	glm-5.3-max	glm-5.3-flash	glm-5.2-max	mimo-v2.5-pro	deepseek-v4.1-flash-max	glm-5.1	kimi-k2.6	mimo-v2.6-flash	deepseek-v4-pro	glm-5
Gemini 4 Argon (High) paid		0.56	0.58	0.59	0.59	0.59	0.60	0.60	0.60	0.61	0.61	0.61	0.62
mimo-v2.6-pro	0.44		0.52	0.53	0.53	0.53	0.54	0.54	0.54	0.55	0.55	0.56	0.56
Kimi K3 (Max)	0.42	0.48		0.51	0.51	0.51	0.51	0.52	0.52	0.53	0.53	0.53	0.54
glm-5.3-max	0.41	0.47	0.49		0.50	0.50	0.51	0.51	0.51	0.52	0.52	0.53	0.54
glm-5.3-flash	0.41	0.47	0.49	0.50		0.50	0.51	0.51	0.51	0.52	0.52	0.53	0.54
glm-5.2-max	0.41	0.47	0.49	0.50	0.50		0.51	0.51	0.51	0.52	0.52	0.53	0.53
mimo-v2.5-pro	0.40	0.46	0.49	0.49	0.49	0.49		0.50	0.50	0.51	0.51	0.52	0.53
deepseek-v4.1-flash-max	0.40	0.46	0.48	0.49	0.49	0.49	0.50		0.50	0.51	0.51	0.52	0.52
glm-5.1	0.40	0.46	0.48	0.49	0.49	0.49	0.50	0.50		0.51	0.51	0.52	0.52
kimi-k2.6	0.39	0.45	0.47	0.48	0.48	0.48	0.49	0.49	0.49		0.50	0.51	0.51
mimo-v2.6-flash	0.39	0.45	0.47	0.48	0.48	0.48	0.49	0.49	0.49	0.50		0.51	0.51
deepseek-v4-pro	0.39	0.44	0.47	0.47	0.47	0.47	0.48	0.48	0.48	0.49	0.49		0.51
glm-5	0.38	0.44	0.46	0.46	0.46	0.47	0.47	0.48	0.48	0.49	0.49	0.49

A usually loses (down to 0.38)Even (0.50)A usually wins (up to 0.62)

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.

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	29.5M	0.8 B	~3 GB at 4-bit · simulate
Qwen3-4B Qwen	11.8M	4 B	~5 GB at 4-bit · simulate
Qwen3-8B Qwen	8.9M	8.2 B	~7 GB at 4-bit · simulate
Qwen3-14B Qwen	8.4M	14.8 B	~12 GB at 4-bit · simulate
Qwen3.5-4B Qwen	7.8M	4.7 B	~5 GB at 4-bit · simulate
Qwen3.5-9B Qwen	7.5M	9.7 B	~8 GB at 4-bit · simulate
gemma-3-1b-it google · license to accept	1.6M	1 B	~3 GB at 4-bit · simulate
NVIDIA-Nemotron-3-Nano-4B-BF16 nvidia	1.6M	4 B	~5 GB at 4-bit · simulate
SmolVLM2-500M-Video-Instruct HuggingFaceTB	1.4M	0.5 B	~2 GB at 4-bit · simulate
MiniCPM5-2B openbmb	1.3M	2.5 B	~4 GB at 4-bit · simulate
DeepSeek-R1-Distill-Qwen-1.5B deepseek-ai	1.3M	1.8 B	~3 GB at 4-bit · simulate
gemma-3-4b-it google · license to accept	1.1M	4.3 B	~5 GB at 4-bit · simulate
SmolLM3-3B HuggingFaceTB	749K	3.1 B	~4 GB at 4-bit · simulate
granite-4.1-3b ibm-granite	648K	3.4 B	~4 GB at 4-bit · simulate
Apertus-v1.5-8B swiss-ai · license to accept	556K	8.9 B	~8 GB at 4-bit · simulate
Olmo-3-7B-Think allenai	489K	7.3 B	~7 GB at 4-bit · simulate
DeepSeek-R1-0528-Qwen3-8B deepseek-ai	445K	8.2 B	~7 GB at 4-bit · simulate
Apertus-8B-Instruct-2509 swiss-ai	434K	8.1 B	~7 GB at 4-bit · simulate
Olmo-3-7B-Instruct allenai	400K	7.3 B	~7 GB at 4-bit · simulate
DeepSeek-R1-Distill-Qwen-7B deepseek-ai	386K	7.6 B	~7 GB at 4-bit · simulate

58 of 145 models in Arena's Coding ranking are free to use. Votes as of 18 Oct 2026. Best free: DeepSeek-V4.1-Flash-20261017 (Max), #12 overall.

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	DeepSeek-V4.1-Flash-20261017 (Max) deepseek	1670 #12 of 145	—	Yes MIT	—	—	—	Lab chat
2	Kimi K3 (Max) moonshot	1655 #15 of 145	2779.9 B 105.8 B active	Yes Kimi K3 license	1686 GB	+1.2% very small loss	does not fit	Lab chat
3	MiMo-V2.6-Flash xiaomi	1633 #18 of 145	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
4	Hy4 preview tencent	1631 #19 of 145	—	Yes Apache 2.0	—	—	—	Lab chat
5	Qwen3.8-Flash-Next alibaba	1629 #20 of 145	180 B 61.6 B active	Yes qwen-community-1.0	110 GB	+1.2% very small loss	does not fit	Lab chat
6	MiMo-V2.6-Pro xiaomi	1626 #21 of 145	1024.2 B 44.9 B active	Yes MIT	622 GB	+1.2% very small loss	does not fit	—
7	GLM-5.3 (Max) zai	1622 #23 of 145	753.3 B 51.2 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
8	DeepSeek-V4.1-Flash-20260910 (Max) deepseek	1619 #25 of 145	—	Yes MIT	—	—	—	Lab chat
9	GLM-5.3-Flash zai	1607 #28 of 145	321.3 B 25.4 B active	Yes MIT	195 GB	+1.2% very small loss	does not fit	Lab chat
10	GLM-5.2 (Max) zai	1604 #29 of 145	753.3 B 51.2 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
11	Qwen3.8-27B alibaba	1592 #30 of 145	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
12	DeepSeek-V4-Flash (High) deepseek	1582 #34 of 145	290.9 B 20.4 B active	Yes MIT	178 GB	+1.2% very small loss	does not fit	Lab chat
13	DeepSeek-V4-Pro (High, 20260813) deepseek	1582 #35 of 145	1598.8 B 75.6 B active	Yes MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
14	Muse Spark Open (Max) meta	1524 #50 of 145	—	Yes muse-model-license	—	—	—	—
15	Kimi K2.6 moonshot	1510 #57 of 145	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
16	GLM-5.1 zai	1508 #58 of 145	753.9 B 51.7 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
17	Hy3 tencent	1507 #59 of 145	298.8 B 24.4 B active	Yes Apache 2.0	184 GB	+1.2% very small loss	does not fit	Lab chat
18	MiniMax-M3 minimax	1482 #65 of 145	427 B 26.8 B active	Yes MiniMax Community License	260 GB	+1.2% very small loss	does not fit	Lab chat
19	MiMo-V2.5-Pro xiaomi	1481 #66 of 145	1023.2 B 43.8 B active	Yes MIT	621 GB	+1.2% very small loss	does not fit	—
20	Kimi K2.7 Code moonshot	1474 #67 of 145	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
21	DeepSeek-V4-Pro Preview (High) deepseek	1465 #69 of 145	1598.8 B 75.6 B active	Yes MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
22	DeepSeek-V4-Pro deepseek	1447 #73 of 145	1598.8 B 75.6 B active	Yes MIT	970 GB	+1.2% very small loss	does not fit	Lab chat
23	MiMo-V2.5 xiaomi	1439 #78 of 145	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
24	beam-1.0-501b-a23b reflection	1437 #79 of 145	—	Yes Apache 2.0	—	—	—	—
25	Kimi K2.5 (Thinking) moonshot	1437 #80 of 145	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
26	GLM-5 zai	1436 #81 of 145	753.9 B 51.8 B active	Yes MIT	458 GB	+1.2% very small loss	does not fit	Lab chat
27	GLM-4.7 zai	1435 #82 of 145	358.3 B 39.1 B active	Yes MIT	221 GB	+1.2% very small loss	does not fit	Lab chat
28	DeepSeek-V4-Flash Preview (High) deepseek	1431 #85 of 145	290.9 B 20.4 B active	Yes MIT	178 GB	+1.2% very small loss	does not fit	Lab chat
29	Inkling thinky	1414 #88 of 145	—	Yes Apache 2.0	—	—	—	Free API · try it
30	Inkling-Small thinky	1410 #89 of 145	—	Yes Apache 2.0	—	—	—	Free API · try it
31	Kimi K2.5 (Instant) moonshot	1406 #91 of 145	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
32	Qwen3.5-397B-A17B alibaba	1400 #94 of 145	403.4 B 17 B active	Yes Apache 2.0	245 GB	+1.2% very small loss	does not fit	Lab chat
33	MiniMax-M2.7 minimax	1399 #95 of 145	228.7 B 11 B active	Yes Modified MIT	141 GB	+1.2% very small loss	does not fit	Lab chat
34	MiniMax-M2.5 minimax	1386 #100 of 145	228.7 B 11 B active	Yes Modified MIT	141 GB	+1.2% very small loss	does not fit	Lab chat
35	MiniMax-M2.1 Preview minimax	1384 #103 of 145	228.7 B 11 B active	Yes MIT	141 GB	+1.2% very small loss	does not fit	Lab chat
36	Gemma 4 31B google	1368 #107 of 145	31.3 B	Yes Apache 2.0	22 GB	+1.8% small loss	33 tok/s	Free API · try it Lab chat
37	DeepSeek-V3.2 (Thinking) deepseek	1363 #108 of 145	685.4 B 51.9 B active	Yes MIT	417 GB	+1.2% very small loss	does not fit	Lab chat
38	Qwen3.5-122B-A10B alibaba	1361 #109 of 145	125.1 B 10 B active	Yes Apache 2.0	77 GB	+1.2% very small loss	does not fit	Lab chat
39	Gemma 4 26B A4B google	1360 #110 of 145	25.8 B 4 B active	Yes Apache 2.0	17 GB	+1.9% small loss	238 tok/s	Free API · try it Lab chat
40	Qwen3.5-27B alibaba	1358 #111 of 145	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
41	Hy3 preview tencent	1357 #113 of 145	298.8 B 24.4 B active	Yes tencent-hunyuan-community	184 GB	+1.2% very small loss	does not fit	Lab chat
42	Muse Glimmer meta	1355 #114 of 145	—	Yes Apache-2.0	—	—	—	—
43	Laguna M.1 poolside	1350 #115 of 145	—	Yes Apache 2.0	—	—	—	—
44	GLM-4.6 zai	1341 #117 of 145	356.8 B 37.6 B active	Yes MIT	220 GB	+1.2% very small loss	does not fit	Lab chat
45	MiMo-V2-Flash (Non-Thinking) xiaomi	1330 #120 of 145	309.8 B 16.5 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
46	DeepSeek-V3.2 deepseek	1326 #122 of 145	—	Yes MIT	—	—	—	Lab chat
47	Kimi K2 Thinking Turbo moonshot	1322 #123 of 145	—	Yes Modified MIT	—	—	—	Lab chat
48	Laguna XS.2 poolside	1303 #124 of 145	—	Yes Apache 2.0	—	—	—	—
49	MiniMax-M2 minimax	1297 #125 of 145	—	Yes Apache 2.0	—	—	—	Lab chat
50	MiMo-V2-Flash (Thinking) xiaomi	1293 #126 of 145	—	Yes MIT	—	—	—	—
51	Qwen3-Coder-480B-A35B-Instruct alibaba	1275 #127 of 145	—	Yes Apache 2.0	—	—	—	Lab chat
52	DeepSeek-V3.2-Exp deepseek	1272 #128 of 145	—	Yes MIT	—	—	—	Lab chat
53	Mistral Medium 3.5 mistral	1264 #129 of 145	—	Yes Modified MIT	—	—	—	—
54	Qwen3.5-35B-A3B alibaba	1254 #132 of 145	—	Yes Apache 2.0	—	—	—	Lab chat
55	Trinity Large Thinking —	1238 #136 of 145	—	Yes Apache 2.0	—	—	—	—
56	Mistral Large 3 mistral	1230 #137 of 145	—	Yes Apache 2.0	—	—	—	—
57	Devstral 2 mistral	1197 #140 of 145	—	Yes Modified MIT	—	—	—	—
58	Granite 4.1 8B ibm	1193 #141 of 145	—	Yes Apache 2.0	—	—	—	—

Top 15 free coding models

DeepSeek-V4.1-Flash-20261017 (Max) 1670 Kimi K3 (Max) 1655 MiMo-V2.6-Flash 1633 Hy4 preview 1631 Qwen3.8-Flash-Next 1629 MiMo-V2.6-Pro 1626 GLM-5.3 (Max) 1622 DeepSeek-V4.1-Flash-20260910 (Max) 1619 GLM-5.3-Flash 1607 GLM-5.2 (Max) 1604 Qwen3.8-27B 1592 DeepSeek-V4-Flash (High) 1582 DeepSeek-V4-Pro (High, 20260813) 1582 Muse Spark Open (Max) 1524 Kimi K2.6 1510

1500160017001800

Arena score, axis starts at 1450. Dashed line: best paid model, Claude Opus 5.5 (Max) (1806).

Score against size

Small and strong Kimi K3 (Max)MiMo-V2.6-FlashQwen3.8-27BGemma 4 26B A4B

1300140015001600

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.

Top free models over the last six months

160017001800

MayJunJulAugSeptOct

DeepSeek-V4.1-Flash-20261017 (Max) 1670Kimi K3 (Max) 1655MiMo-V2.6-Flash 1633Hy4 preview 1631Qwen3.8-Flash-Next 1629Best paid model

Arena score from 20 Apr 2026 to 18 Oct 2026.

Who beats whom

Model A ↓ · Model B →	Claude Opus 5.5 (Max) (paid)	DeepSeek-V4.1-Flash-20261017 (Max)	Kimi K3 (Max)	MiMo-V2.6-Flash	Hy4 preview	Qwen3.8-Flash-Next	MiMo-V2.6-Pro	GLM-5.3 (Max)	DeepSeek-V4.1-Flash-20260910 (Max)	GLM-5.3-Flash	GLM-5.2 (Max)	Qwen3.8-27B	DeepSeek-V4-Flash (High)
Claude Opus 5.5 (Max) paid		0.69	0.70	0.73	0.73	0.73	0.74	0.74	0.75	0.76	0.76	0.77	0.78
DeepSeek-V4.1-Flash-20261017 (Max)	0.31		0.52	0.55	0.56	0.56	0.56	0.57	0.57	0.59	0.59	0.61	0.62
Kimi K3 (Max)	0.30	0.48		0.53	0.53	0.54	0.54	0.55	0.55	0.57	0.57	0.59	0.60
MiMo-V2.6-Flash	0.27	0.45	0.47		0.50	0.51	0.51	0.52	0.52	0.54	0.54	0.56	0.57
Hy4 preview	0.27	0.44	0.47	0.50		0.50	0.51	0.51	0.52	0.53	0.54	0.56	0.57
Qwen3.8-Flash-Next	0.27	0.44	0.46	0.49	0.50		0.50	0.51	0.51	0.53	0.54	0.55	0.57
MiMo-V2.6-Pro	0.26	0.44	0.46	0.49	0.49	0.50		0.51	0.51	0.53	0.53	0.55	0.56
GLM-5.3 (Max)	0.26	0.43	0.45	0.48	0.49	0.49	0.49		0.50	0.52	0.53	0.54	0.56
DeepSeek-V4.1-Flash-20260910 (Max)	0.25	0.43	0.45	0.48	0.48	0.49	0.49	0.50		0.52	0.52	0.54	0.55
GLM-5.3-Flash	0.24	0.41	0.43	0.46	0.47	0.47	0.47	0.48	0.48		0.50	0.52	0.54
GLM-5.2 (Max)	0.24	0.41	0.43	0.46	0.46	0.46	0.47	0.47	0.48	0.50		0.52	0.53
Qwen3.8-27B	0.23	0.39	0.41	0.44	0.44	0.45	0.45	0.46	0.46	0.48	0.48		0.51
DeepSeek-V4-Flash (High)	0.22	0.38	0.40	0.43	0.43	0.43	0.44	0.44	0.45	0.46	0.47	0.49

A usually loses (down to 0.22)Even (0.50)A usually wins (up to 0.78)

49 of 165 models in Arena's Vision ranking are free to use. Votes as of 9 Oct 2026. Best free: glm-5.3-flash, #23 overall.

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	glm-5.3-flash zai	1296 #23 of 165	321.3 B 25.4 B active	Yes MIT	195 GB	+1.2% very small loss	does not fit	Lab chat
2	kimi-k2.6 moonshot	1283 #33 of 165	1026.9 B 33.3 B active	Yes Modified MIT	623 GB	+1.2% very small loss	does not fit	Lab chat
3	gemma-4-31b google	1277 #38 of 165	31.3 B	Yes Apache 2.0	22 GB	+1.8% small loss	33 tok/s	Free API · try it Lab chat
4	deepseek-v4.1-flash-max deepseek	1277 #39 of 165	763.2 B 228.1 B active	Yes MIT	464 GB	+1.2% very small loss	does not fit	Lab chat
5	qwen3.8-27b alibaba	1271 #42 of 165	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
6	kimi-k2.5-thinking moonshot	1269 #44 of 165	—	Yes Modified MIT	—	—	—	Lab chat
7	mimo-v2.6-pro xiaomi	1264 #50 of 165	1024.2 B 44.9 B active	Yes MIT	622 GB	+1.2% very small loss	does not fit	—
8	qwen3.5-397b-a17b alibaba	1263 #54 of 165	403.4 B 17 B active	Yes Apache 2.0	245 GB	+1.2% very small loss	does not fit	Lab chat
9	gemma-4-26b-a4b google	1260 #55 of 165	25.8 B 4 B active	Yes Apache 2.0	17 GB	+1.9% small loss	238 tok/s	Free API · try it Lab chat
10	mimo-v2.6-flash xiaomi	1259 #56 of 165	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
11	kimi-k2.5-instant moonshot	1253 #59 of 165	—	Yes Modified MIT	—	—	—	Lab chat
12	minimax-m3 minimax	1253 #60 of 165	427 B 26.8 B active	Yes MiniMax Community License	260 GB	+1.2% very small loss	does not fit	Lab chat
13	mimo-v2.5 xiaomi	1247 #65 of 165	310.8 B 17.4 B active	Yes MIT	189 GB	+1.2% very small loss	does not fit	—
14	qwen3-vl-235b-a22b-instruct alibaba	1247 #66 of 165	235.7 B 22 B active	Yes Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
15	qwen3.5-122b-a10b alibaba	1245 #70 of 165	125.1 B 10 B active	Yes Apache 2.0	77 GB	+1.2% very small loss	does not fit	Lab chat
16	qwen3.5-27b alibaba	1241 #72 of 165	27.8 B	Yes Apache 2.0	18 GB	+1.9% small loss	41 tok/s	Lab chat
17	Inkling Small thinky	1235 #75 of 165	—	Yes Apache 2.0	—	—	—	Free API · try it
18	mistral-medium-3.5 mistral	1223 #82 of 165	—	Yes Modified MIT	—	—	—	—
19	mistral-large-3 mistral	1221 #83 of 165	—	Yes Apache 2.0	—	—	—	—
20	qwen3-vl-235b-a22b-thinking alibaba	1207 #90 of 165	235.7 B 22 B active	Yes Apache 2.0	145 GB	+1.2% very small loss	does not fit	Lab chat
21	step-3 stepfun	1176 #104 of 165	321 B 18.7 B active	Yes Apache 2.0	228 GB	+1.2% very small loss	does not fit	—
22	gemma-3-27b-it google	1164 #110 of 165	27.4 B	Yes Gemma	18 GB	+1.9% small loss	40 tok/s	Lab chat
23	glm-4.6v zai	1164 #111 of 165	107.7 B 14.3 B active	Yes MIT	67 GB	+1.2% very small loss	12 tok/s GPU + RAM	Lab chat
24	glm-4.5v zai	1154 #116 of 165	107.7 B 14.3 B active	Yes MIT	67 GB	+1.2% very small loss	12 tok/s GPU + RAM	Lab chat
25	qwen2.5-vl-32b-instruct alibaba	1152 #117 of 165	33.5 B	Yes Apache 2.0	23 GB	+1.8% small loss	31 tok/s	Lab chat
26	mistral-small-2506 mistral	1142 #119 of 165	—	Yes Apache 2.0	—	—	—	—
27	llama-4-maverick-17b-128e-instruct meta	1142 #120 of 165	401.6 B	Yes Llama 4	246 GB	+1.2% very small loss	does not fit	—
28	mistral-small-3.1-24b-instruct-2503 mistral	1136 #123 of 165	—	Yes Apache 2.0	—	—	—	—
29	llama-4-scout-17b-16e-instruct meta	1118 #127 of 165	—	Yes Llama	—	—	—	—
30	qwen2.5-vl-72b-instruct alibaba	1107 #128 of 165	—	Yes Qwen	—	—	—	Lab chat
31	pixtral-large-2411 mistral	1089 #132 of 165	—	Yes MRL	—	—	—	—
32	molmo-2-8b allenai	1081 #134 of 165	—	Yes Apache 2.0	—	—	—	—
33	qwen2-vl-72b alibaba	1047 #139 of 165	—	Yes Qwen	—	—	—	Lab chat
34	molmo-72b-0924 allenai	1012 #143 of 165	—	Yes Apache 2.0	—	—	—	—
35	pixtral-12b-2409 mistral	1009 #144 of 165	—	Yes Apache 2.0	—	—	—	—
36	internvl2-26b —	1005 #146 of 165	—	Yes MIT	—	—	—	—
37	llama-3.2-vision-90b-instruct meta	1000 #147 of 165	—	Yes Llama 3.2	—	—	—	—
38	c4ai-aya-vision-32b cohere	996 #149 of 165	—	Yes CC-BY-NC-4.0	—	—	—	—
39	qwen2-vl-7b-instruct alibaba	990 #151 of 165	—	Yes Apache 2.0	—	—	—	Lab chat
40	llama-3.2-vision-11b-instruct meta	966 #155 of 165	—	Yes Llama 3.2	—	—	—	—
41	molmo-7b-d-0924 allenai	959 #156 of 165	—	Yes Apache 2.0	—	—	—	—
42	internvl2-4b —	941 #158 of 165	—	Yes MIT	—	—	—	—
43	nvila-internal-15b-v1 nvidia	940 #159 of 165	—	Yes -	—	—	—	Lab chat
44	llava-v1.6-34b —	936 #160 of 165	—	Yes Apache 2.0	—	—	—	—
45	cogvlm2-llama3-chat-19b zai	928 #161 of 165	—	Yes CogVLM2	—	—	—	Lab chat
46	llava-onevision-qwen2-72b-ov —	923 #162 of 165	—	Yes Apache 2.0	—	—	—	—
47	minicpm-v-2_6 —	912 #163 of 165	—	Yes Apache 2.0	—	—	—	—
48	phi-3.5-vision-instruct microsoft	851 #164 of 165	—	Yes MIT	—	—	—	—
49	phi-3-vision-128k-instruct microsoft	812 #165 of 165	—	Yes MIT	—	—	—	—

Top 15 free vision models

glm-5.3-flash 1296 kimi-k2.6 1283 gemma-4-31b 1277 deepseek-v4.1-flash-max 1277 qwen3.8-27b 1271 kimi-k2.5-thinking 1269 mimo-v2.6-pro 1264 qwen3.5-397b-a17b 1263 gemma-4-26b-a4b 1260 mimo-v2.6-flash 1259 kimi-k2.5-instant 1253 minimax-m3 1253 mimo-v2.5 1247 qwen3-vl-235b-a22b-instruct 1247 qwen3.5-122b-a10b 1245

1150120012501300

Arena score, axis starts at 1150. Dashed line: best paid model, claude-fable-5-high (1324).

Score against size

Small and strong glm-5.3-flashgemma-4-31bgemma-4-26b-a4b

1150120012501300

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.

Top free models over the last six months

12501275130013251350

MayJunJulAugSeptOct

glm-5.3-flash 1296kimi-k2.6 1283gemma-4-31b 1277deepseek-v4.1-flash-max 1277qwen3.8-27b 1271Best paid model

Arena score from 10 Apr 2026 to 9 Oct 2026.

Who beats whom

Model A ↓ · Model B →	claude-fable-5-high (paid)	glm-5.3-flash	kimi-k2.6	gemma-4-31b	deepseek-v4.1-flash-max	qwen3.8-27b	kimi-k2.5-thinking	mimo-v2.6-pro	qwen3.5-397b-a17b	gemma-4-26b-a4b	mimo-v2.6-flash	kimi-k2.5-instant	minimax-m3
claude-fable-5-high paid		0.54	0.56	0.57	0.57	0.58	0.58	0.59	0.59	0.59	0.59	0.60	0.60
glm-5.3-flash	0.46		0.52	0.53	0.53	0.54	0.54	0.55	0.55	0.55	0.55	0.56	0.56
kimi-k2.6	0.44	0.48		0.51	0.51	0.52	0.52	0.53	0.53	0.53	0.53	0.54	0.54
gemma-4-31b	0.43	0.47	0.49		0.50	0.51	0.51	0.52	0.52	0.52	0.53	0.53	0.53
deepseek-v4.1-flash-max	0.43	0.47	0.49	0.50		0.51	0.51	0.52	0.52	0.52	0.53	0.53	0.53
qwen3.8-27b	0.42	0.46	0.48	0.49	0.49		0.50	0.51	0.51	0.52	0.52	0.53	0.53
kimi-k2.5-thinking	0.42	0.46	0.48	0.49	0.49	0.50		0.51	0.51	0.51	0.51	0.52	0.52
mimo-v2.6-pro	0.41	0.45	0.47	0.48	0.48	0.49	0.49		0.50	0.51	0.51	0.52	0.52
qwen3.5-397b-a17b	0.41	0.45	0.47	0.48	0.48	0.49	0.49	0.50		0.50	0.51	0.51	0.51
gemma-4-26b-a4b	0.41	0.45	0.47	0.48	0.48	0.48	0.49	0.49	0.50		0.50	0.51	0.51
mimo-v2.6-flash	0.41	0.45	0.47	0.47	0.47	0.48	0.49	0.49	0.49	0.50		0.51	0.51
kimi-k2.5-instant	0.40	0.44	0.46	0.47	0.47	0.47	0.48	0.48	0.49	0.49	0.49		0.50
minimax-m3	0.40	0.44	0.46	0.47	0.47	0.47	0.48	0.48	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)

Small models, no Arena votes yet

Model	Downloads/month	Size	Run at home
Qwen3.5-4B Qwen	7.8M	4.7 B	~5 GB at 4-bit · simulate
Qwen3.5-9B Qwen	7.5M	9.7 B	~8 GB at 4-bit · simulate
SmolVLM2-500M-Video-Instruct HuggingFaceTB	1.4M	0.5 B	~2 GB at 4-bit · simulate
gemma-3-4b-it google · license to accept	1.1M	4.3 B	~5 GB at 4-bit · simulate
Apertus-v1.5-8B swiss-ai · license to accept	556K	8.9 B	~8 GB at 4-bit · simulate
gemma-3-12b-it google · license to accept	361K	12.2 B	~10 GB at 4-bit · simulate
MiniCPM-V-4.6 openbmb	204K	1.3 B	~3 GB at 4-bit · simulate
granite-vision-4.1-4b ibm-granite	160K	4 B	~5 GB at 4-bit · simulate
GLM-4.1V-9B-Thinking zai-org	120K	10.3 B	~9 GB at 4-bit · simulate
gemma-3n-E2B-it google · license to accept	119K	5.4 B	~6 GB at 4-bit · simulate
SmolVLM2-2.2B-Instruct HuggingFaceTB	119K	2.2 B	~3 GB at 4-bit · simulate
SmolVLM-500M-Instruct HuggingFaceTB	108K	0.5 B	~2 GB at 4-bit · simulate
GLM-4.6V-Flash zai-org	99K	10.3 B	~9 GB at 4-bit · simulate
MiniCPM-V-4_5 openbmb	86K	8.7 B	~8 GB at 4-bit · simulate
Llama-3.1-Nemotron-Nano-VL-8B-V1 nvidia	80K	8.7 B	~8 GB at 4-bit · simulate
MiniCPM-V-4.6-Thinking openbmb	70K	1.3 B	~3 GB at 4-bit · simulate
North-Micro-Vision-Instruct CohereLabs	46K	2.5 B	~4 GB at 4-bit · simulate
Molmo2-8B allenai	45K	8.7 B	~8 GB at 4-bit · simulate
MiniCPM-V-4 openbmb	40K	4.1 B	~5 GB at 4-bit · simulate
LFM2.5-VL-1.6B LiquidAI	30K	1.6 B	~3 GB at 4-bit · simulate

19 of 82 models in Arena's Image ranking are free to use. Votes as of 7 Oct 2026. Best free: qwen-image-2.1, #19 overall.

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	qwen-image-2.1 alibaba	1223 #19 of 82	7.1 B	Yes qwen-research	7.9 GB	—	fits	Lab chat
2	ideogram-4.0-quality ideogram	1205 #20 of 82	—	Yes Ideogram Open Model	—	—	—	—
3	Cosmos3-Super-Text2Image (Agentic) nvidia	1165 #27 of 82	—	Yes OpenMDW-1.1	—	—	—	Lab chat
4	Cosmos3-Super-Text2Image nvidia	1152 #33 of 82	64.6 B	Yes OpenMDW-1.1	43 GB	—	with RAM GPU + RAM	Lab chat
5	hunyuan-image-3.0 tencent	1150 #35 of 82	83 B	Yes tencent-hunyuan-community	54 GB	—	with RAM GPU + RAM	Lab chat
6	flux-2-dev bfl	1146 #38 of 82	32.2 B	Yes flux-non-commercial-license	23 GB	—	fits	—
7	qwen-image-2512 alibaba	1125 #44 of 82	20.4 B	Yes Apache 2.0	16 GB	—	fits	Lab chat
8	hidream-o1-image hidream	1116 #47 of 82	8.8 B	Yes MIT	8.9 GB	—	fits	—
9	krea-2-turbo krea	1113 #52 of 82	12.8 B	Yes krea-2-community-license	11 GB	—	fits	—
10	z-image-turbo alibaba	1083 #59 of 82	6.2 B	Yes Apache 2.0	7.4 GB	—	fits	Lab chat
11	flux-2-klein-9b bfl	1070 #62 of 82	9.1 B	Yes flux-non-commercial-license	9.1 GB	—	fits	—
12	qwen-image-prompt-extend alibaba	1060 #63 of 82	—	Yes Apache 2.0	—	—	—	Lab chat
13	qwen-image alibaba	1057 #66 of 82	20.4 B	Yes Apache 2.0	16 GB	—	fits	Lab chat
14	flux-2-klein-4b bfl	1030 #70 of 82	3.9 B	Yes Apache 2.0	6.0 GB	—	fits	—
15	glm-image zai	1012 #76 of 82	6.9 B	Yes MIT	7.8 GB	—	fits	Lab chat
16	flux-1-dev-fp8 bfl	969 #78 of 82	—	Yes Open	—	—	—	—
17	flux-1-kontext-dev bfl	941 #80 of 82	11.9 B	Yes flux-1-dev-non-commercial-license	11 GB	—	fits	—
18	stable-diffusion-v35-large —	938 #81 of 82	8.1 B	Yes Open	8.5 GB	—	fits	—
19	bagel bytedance	898 #82 of 82	—	Yes Apache 2.0	—	—	—	—

Size: parameters (and active parameters for mixture-of-experts). 4-bit needs: GPU memory at Q4_K_M plus about 3 GB for encoders. Speed: estimate on one RTX 4090 (24 GB) with 64 GB of DDR5, at 4-bit. Simulate your own hardware.

Top 15 free image models

qwen-image-2.1 1223 ideogram-4.0-quality 1205 Cosmos3-Super-Text2Image (Agentic) 1165 Cosmos3-Super-Text2Image 1152 hunyuan-image-3.0 1150 flux-2-dev 1146 qwen-image-2512 1125 hidream-o1-image 1116 krea-2-turbo 1113 z-image-turbo 1083 flux-2-klein-9b 1070 qwen-image-prompt-extend 1060 qwen-image 1057 flux-2-klein-4b 1030 glm-image 1012

10001100120013001400

Arena score, axis starts at 950. Dashed line: best paid model, gpt-image-2.5-sunburst (1425).

Score against size

Small and strong qwen-image-2.1z-image-turboflux-2-klein-4b

100011001200

10B100B

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.

Top free models over the last six months

12001400

MayJunJulAugSeptOct

qwen-image-2.1 1223ideogram-4.0-quality 1205Cosmos3-Super-Text2Image (Agentic) 1165Cosmos3-Super-Text2Image 1152hunyuan-image-3.0 1150Best paid model

Arena score from 10 Apr 2026 to 7 Oct 2026.

Who beats whom

Model A ↓ · Model B →	gpt-image-2.5-sunburst (paid)	qwen-image-2.1	ideogram-4.0-quality	Cosmos3-Super-Text2Image (Agentic)	Cosmos3-Super-Text2Image	hunyuan-image-3.0	flux-2-dev	qwen-image-2512	hidream-o1-image	krea-2-turbo	z-image-turbo	flux-2-klein-9b	qwen-image-prompt-extend
gpt-image-2.5-sunburst paid		0.76	0.78	0.82	0.83	0.83	0.83	0.85	0.86	0.86	0.88	0.89	0.89
qwen-image-2.1	0.24		0.53	0.58	0.60	0.60	0.61	0.64	0.65	0.65	0.69	0.71	0.72
ideogram-4.0-quality	0.22	0.47		0.56	0.58	0.58	0.58	0.61	0.63	0.63	0.67	0.69	0.70
Cosmos3-Super-Text2Image (Agentic)	0.18	0.42	0.44		0.52	0.52	0.53	0.56	0.57	0.57	0.62	0.63	0.65
Cosmos3-Super-Text2Image	0.17	0.40	0.42	0.48		0.50	0.51	0.54	0.55	0.56	0.60	0.62	0.63
hunyuan-image-3.0	0.17	0.40	0.42	0.48	0.50		0.51	0.54	0.55	0.55	0.60	0.61	0.63
flux-2-dev	0.17	0.39	0.42	0.47	0.49	0.49		0.53	0.54	0.55	0.59	0.61	0.62
qwen-image-2512	0.15	0.36	0.39	0.44	0.46	0.46	0.47		0.51	0.52	0.56	0.58	0.59
hidream-o1-image	0.14	0.35	0.37	0.43	0.45	0.45	0.46	0.49		0.50	0.55	0.57	0.58
krea-2-turbo	0.14	0.35	0.37	0.43	0.44	0.45	0.45	0.48	0.50		0.54	0.56	0.58
z-image-turbo	0.12	0.31	0.33	0.38	0.40	0.40	0.41	0.44	0.45	0.46		0.52	0.53
flux-2-klein-9b	0.11	0.29	0.31	0.37	0.38	0.39	0.39	0.42	0.43	0.44	0.48		0.51
qwen-image-prompt-extend	0.11	0.28	0.30	0.35	0.37	0.37	0.38	0.41	0.42	0.42	0.47	0.49

A usually loses (down to 0.11)Even (0.50)A usually wins (up to 0.89)

7 of 48 models in Arena's Video ranking are free to use. Votes as of 22 Sept 2026. Best free: minimax-h3, #8 overall.

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	minimax-h3 minimax	1460 #8 of 48	33.1 B	Yes minimax-h3-community-license-agreement	24 GB	—	fits	Lab chat
2	kandinsky-5.0-t2v-pro kandinsky	1175 #37 of 48	19.3 B	Yes MIT	15 GB	—	fits	—
3	hunyuan-video-1.5 tencent	1169 #38 of 48	8.3 B	Yes tencent-hunyuan-community	8.6 GB	—	fits	Lab chat
4	ltx-2-19b —	1154 #41 of 48	18.9 B	Yes ltx-2-community-license-agreement	15 GB	—	fits	—
5	wan-v2.2-a14b alibaba	1132 #42 of 48	27 B	Yes Apache 2.0	20 GB	—	fits	Lab chat
6	kandinsky-5.0-t2v-lite kandinsky	1113 #44 of 48	2 B	Yes MIT	4.8 GB	—	fits	—
7	mochi-v1 genmo	1007 #48 of 48	10 B	Yes Apache 2.0	9.7 GB	—	fits	—

Score against size

Small and strong minimax-h3kandinsky-5.0-t2v-prokandinsky-5.0-t2v-lite

100012001400

1B10B100B

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.

Top free models over the last six months

12001400

MayJunJulAugSept

minimax-h3 1460kandinsky-5.0-t2v-pro 1175hunyuan-video-1.5 1169ltx-2-19b 1154wan-v2.2-a14b 1132Best paid model

Arena score from 7 Apr 2026 to 22 Sept 2026.

Who beats whom

Model A ↓ · Model B →	gemini-omni-1.1-flash (paid)	minimax-h3	kandinsky-5.0-t2v-pro	hunyuan-video-1.5	ltx-2-19b	wan-v2.2-a14b	kandinsky-5.0-t2v-lite	mochi-v1
gemini-omni-1.1-flash paid		0.58	0.88	0.88	0.89	0.90	0.91	0.95
minimax-h3	0.42		0.84	0.84	0.85	0.87	0.88	0.93
kandinsky-5.0-t2v-pro	0.12	0.16		0.51	0.53	0.56	0.59	0.72
hunyuan-video-1.5	0.12	0.16	0.49		0.52	0.55	0.58	0.72
ltx-2-19b	0.11	0.15	0.47	0.48		0.53	0.56	0.70
wan-v2.2-a14b	0.10	0.13	0.44	0.45	0.47		0.53	0.67
kandinsky-5.0-t2v-lite	0.09	0.12	0.41	0.42	0.44	0.47		0.65
mochi-v1	0.05	0.07	0.28	0.28	0.30	0.33	0.35

A usually loses (down to 0.05)Even (0.50)A usually wins (up to 0.95)

10 of 59 models in Arena's Image edit ranking are free to use. Votes as of 6 Oct 2026. Best free: qwen-image-2.1, #18 overall.

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	qwen-image-2.1 alibaba	1369 #18 of 59	7.1 B	Yes qwen-research	7.9 GB	—	fits	Lab chat
2	hunyuan-image-3.0-instruct tencent	1303 #26 of 59	83 B	Yes tencent-hunyuan-community	54 GB	—	with RAM GPU + RAM	Lab chat
3	qwen-image-edit alibaba	1242 #37 of 59	20.4 B	Yes Apache 2.0	16 GB	—	fits	Lab chat
4	qwen-image-edit-2511 alibaba	1234 #39 of 59	20.4 B	Yes Apache 2.0	16 GB	—	fits	Lab chat
5	flux-2-klein-9b bfl	1225 #42 of 59	9.1 B	Yes flux-non-commercial-license	9.1 GB	—	fits	—
6	flux-2-dev bfl	1225 #43 of 59	32.2 B	Yes flux-non-commercial-license	23 GB	—	fits	—
7	flux-2-klein-4b bfl	1187 #49 of 59	3.9 B	Yes Apache 2.0	6.0 GB	—	fits	—
8	flux-1-kontext-dev bfl	1149 #53 of 59	11.9 B	Yes flux-1-dev-non-commercial-license	11 GB	—	fits	—
9	bagel bytedance	1027 #58 of 59	—	Yes Apache 2.0	—	—	—	—
10	step1x-edit stepfun	998 #59 of 59	—	Yes Apache 2.0	—	—	—	—

Top 15 free image edit models

qwen-image-2.1 1369 hunyuan-image-3.0-instruct 1303 qwen-image-edit 1242 qwen-image-edit-2511 1234 flux-2-klein-9b 1225 flux-2-dev 1225 flux-2-klein-4b 1187 flux-1-kontext-dev 1149 bagel 1027 step1x-edit 998

100012001400

Arena score, axis starts at 900. Dashed line: best paid model, gpt-image-2.5-sunburst (1524).

Score against size

Small and strong qwen-image-2.1flux-2-klein-4b

120013001400

10B100B

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.

Top free models over the last six months

1200130014001500

MayJunJulAugSeptOct

qwen-image-2.1 1369hunyuan-image-3.0-instruct 1303qwen-image-edit 1242qwen-image-edit-2511 1234flux-2-klein-9b 1225Best paid model

Arena score from 10 Apr 2026 to 6 Oct 2026.

Who beats whom

Model A ↓ · Model B →	gpt-image-2.5-sunburst (paid)	qwen-image-2.1	hunyuan-image-3.0-instruct	qwen-image-edit	qwen-image-edit-2511	flux-2-klein-9b	flux-2-dev	flux-2-klein-4b	flux-1-kontext-dev	bagel	step1x-edit
gpt-image-2.5-sunburst paid		0.71	0.78	0.84	0.84	0.85	0.85	0.87	0.90	0.95	0.95
qwen-image-2.1	0.29		0.59	0.68	0.69	0.70	0.70	0.74	0.78	0.88	0.89
hunyuan-image-3.0-instruct	0.22	0.41		0.59	0.60	0.61	0.61	0.66	0.71	0.83	0.85
qwen-image-edit	0.16	0.32	0.41		0.51	0.52	0.52	0.58	0.63	0.78	0.80
qwen-image-edit-2511	0.16	0.31	0.40	0.49		0.51	0.51	0.57	0.62	0.77	0.80
flux-2-klein-9b	0.15	0.30	0.39	0.48	0.49		0.50	0.55	0.61	0.76	0.79
flux-2-dev	0.15	0.30	0.39	0.48	0.49	0.50		0.55	0.61	0.76	0.79
flux-2-klein-4b	0.13	0.26	0.34	0.42	0.43	0.45	0.45		0.55	0.72	0.75
flux-1-kontext-dev	0.10	0.22	0.29	0.37	0.38	0.39	0.39	0.45		0.67	0.70
bagel	0.05	0.12	0.17	0.22	0.23	0.24	0.24	0.28	0.33		0.54
step1x-edit	0.05	0.11	0.15	0.20	0.20	0.21	0.21	0.25	0.30	0.46

A usually loses (down to 0.05)Even (0.50)A usually wins (up to 0.95)

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
17 Oct 2026	DeepSeek-V4.1-Flash-1017 deepseek-ai	Chat, Vision	763.2 B 228.1 B active	511 GB	Download
16 Oct 2026	NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 nvidia	Chat	4 B	5 GB	Download
15 Oct 2026	Hy-Memory-30B-A3B tencent	Chat	30.1 B 3 B active	22 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	Ling-3.1-flash inclusionAI	Chat	572.1 B 33.1 B active	383 GB	Download
13 Oct 2026	flux-3-image-klein black-forest-labs	Image edit	5.5 B	6 GB	Download
12 Oct 2026	NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 nvidia	Chat, Vision	124.5 B 12 B active	85 GB	Free API · try itDownload
12 Oct 2026	LFM2.5-24B-A2B LiquidAI	Chat	24 B 2 B active	18 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	NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB nvidia	Chat, Vision	22.4 B 9 B active	17 GB	Download
7 Oct 2026	MiniCPM-V-4.7-35B-A3B openbmb	Chat, Vision	35.2 B 3 B active	25 GB	Download
6 Oct 2026	MiniCPM-V-4.7-1B openbmb	Chat, Vision	1.6 B	3 GB	Download
5 Oct 2026	d1-omni-600M LiquidAI	Chat, Vision	0.6 B	2 GB	Download
5 Oct 2026	d1-3B LiquidAI	Chat, Vision	3.1 B	4 GB	Download
29 Sept 2026	AesCode-32B microsoft	Chat, Vision, Coding	33.4 B	24 GB	Download
29 Sept 2026	AesCode-8B microsoft	Chat, Vision, Coding	8.8 B	8 GB	Download
17 Sept 2026	Ming-Image-0.1-Design inclusionAI	Image	6.2 B	6 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
6 Sept 2026	LLaDA-UI inclusionAI	Chat, Vision	16.9 B	13 GB	Download
2 Sept 2026	tiny-aya-l2-thinker CohereLabs	Chat	3.4 B	4 GB	Download
27 Aug 2026	Qwen-Drive-1.0-4B Qwen	Chat, Vision	4.5 B	5 GB	Download
27 Aug 2026	ContextPilot-14B tencent	Chat	14.8 B	12 GB	Download
free API	XingChen: Xing4.0 29B A4B xingchen-agi	Chat	—	—	Free API · try it
free API	Dots Studio: Dots3-Note Preview dots-studio	Chat	—	—	Free API · try it
free API	LiquidAI: LFM2.5-2.6B liquid	Chat	—	—	Free API · try it
free API	NVIDIA: Nemotron 3.5 Lightning nvidia	Chat	—	—	Free API · try it
free API	Poolside: Laguna S 2.1 poolside	Chat	—	—	Free API · try it
free API	Poolside: Laguna XS 2.1 poolside	Chat	—	—	Free API · try it
free API	Cohere: North Mini Code cohere	Chat	—	—	Free API · try it
free API	NVIDIA: Nemotron 3 Ultra nvidia	Chat	—	—	Free API · try it
free API	NVIDIA: Nemotron 3 Nano Omni nvidia	Chat	—	—	Free API · try it
free API	NVIDIA: Nemotron 3 Super 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.

The best free AI models, ranked by real votes

Full free ranking

Top 15 free chat models

Score against size

Top free models over the last six months

Who beats whom

Small models, no Arena votes yet

Full free ranking

Top 15 free coding models

Score against size

Top free models over the last six months

Who beats whom

Full free ranking

Top 15 free vision models

Score against size

Top free models over the last six months

Who beats whom

Small models, no Arena votes yet

Full free ranking

Top 15 free image models

Score against size

Top free models over the last six months

Who beats whom

Full free ranking

Top 15 free video models

Score against size

Top free models over the last six months

Who beats whom

Full free ranking

Top 15 free image edit models

Score against size

Top free models over the last six months

Who beats whom

New, waiting for Arena votes

Frequently asked questions