LLM Bench — Независимый трекер открытых языковых моделей

Рейтинг моделей

Данные: Artificial Analysis (Intelligence Index) · TIGER-Lab MMLU-Pro · Arena AI

#	Модель	Intelligence ↓	Coding	Speed t/s	TTFT	$/1M out	License
1	Gemini 4 Carbon HIGH Proprietary REASONING Google	59.4	—	95	140с	$20.0	Proprietary
2	Claude Opus 5.5 MAX, DEFAULT F Proprietary REASONING Anthropic	58.4	—	96	721с	$20.0	Proprietary
3	Gemini 4 Carbon MED Proprietary REASONING Google	57.7	—	—	—	$20.0	Proprietary
4	Claude Opus 5.5 MAX Proprietary REASONING Anthropic	57.0	—	83	129с	$20.0	Proprietary
5	GPT-6 Astra MAX Proprietary REASONING OpenAI	56.9	—	74	370с	$50.0	Proprietary
6	GPT-6 Astra MAX Proprietary REASONING OpenAI	55.3	—	70	195с	$50.0	Proprietary
7	GPT-6 Astra HIGH Proprietary REASONING OpenAI	54.8	—	63	96с	$50.0	Proprietary
8	GPT-6.1 Sol MAX Proprietary REASONING OpenAI	54.5	—	91	333с	$10.0	Proprietary
9	Claude Sonnet 5.5 MAX, DEFAULT F Proprietary REASONING Anthropic	54.2	—	138	524с	$10.0	Proprietary
10	GPT-6.1 Sol MAX Proprietary REASONING OpenAI	52.3	—	75	153с	$10.0	Proprietary
11	Claude Opus 5.5 HIGH Proprietary REASONING Anthropic	51.7	—	80	48с	$20.0	Proprietary
12	GPT-6 Astra MED Proprietary REASONING OpenAI	51.5	—	57	6.3с	$50.0	Proprietary
13	GPT-6.1 Terra MAX Proprietary REASONING OpenAI	51.0	—	107	264с	$5.0	Proprietary
14	Gemini 4 Argon HIGH Proprietary REASONING Google	50.4	—	83	40с	$20.0	Proprietary
15	GPT-6.1 Sol HIGH Proprietary REASONING OpenAI	50.2	—	67	80с	$10.0	Proprietary
16	Claude Fable 5.1 MAX Proprietary REASONING Anthropic	49.7	—	63	76с	$50.0	Proprietary
17	Muse Wonder Full MAX Proprietary REASONING Meta	49.7	—	—	—	—	Proprietary
18	Claude Sonnet 5.5 MAX Proprietary REASONING Anthropic	49.6	—	109	33с	$10.0	Proprietary
19	Claude Fable 5.1 MAX, DEFAULT F Proprietary REASONING Anthropic	49.0	—	71	267с	$50.0	Proprietary
20	Claude Opus 5.5 MED Proprietary REASONING Anthropic	48.1	—	80	21с	$20.0	Proprietary
21	Claude Fable 5.1 HIGH Proprietary REASONING Anthropic	47.0	—	58	18с	$50.0	Proprietary
22	GPT-6.1 Sol MED Proprietary REASONING OpenAI	46.1	—	63	8.1с	$10.0	Proprietary
23	Kimi K3.1 Pro Proprietary REASONING Kimi	46.0	—	31	3.4с	$10.0	Proprietary
24	GPT-6.1 Terra MAX Proprietary REASONING OpenAI	45.0	—	99	71с	$5.0	Proprietary
25	Muse Wonder MAX Proprietary REASONING Meta	43.3	—	192	26с	$20.0	Proprietary
26	Claude Fable 5.1 MED Proprietary REASONING Anthropic	43.0	—	58	13с	$50.0	Proprietary
27	GPT-6 Astra LOW Proprietary REASONING OpenAI	43.0	—	57	2.7с	$50.0	Proprietary
28	GPT-6.1 Terra HIGH Proprietary REASONING OpenAI	42.3	—	95	21с	$5.0	Proprietary
29	Claude Sonnet 5.5 HIGH Proprietary REASONING Anthropic	41.1	—	108	12с	$10.0	Proprietary
30	Claude Fable 5.1 LOW Proprietary REASONING Anthropic	38.5	—	57	3.1с	$50.0	Proprietary
31	GPT-6.1 Terra MED Proprietary REASONING OpenAI	38.1	—	91	6.3с	$5.0	Proprietary
32	Claude Haiku 5.5 MAX Proprietary REASONING Anthropic	36.2	—	236	358с	$0.50	Proprietary
33	Grok 4.7 HIGH Proprietary REASONING SpaceXAI	35.0	—	76	57с	$6.0	Proprietary
34	Claude Opus 5.5 LOW Proprietary REASONING Anthropic	34.5	—	79	11с	$20.0	Proprietary
35	GPT-6.1 Sol LOW Proprietary REASONING OpenAI	34.5	—	61	2.4с	$10.0	Proprietary
36	Claude Haiku 5.5 MAX Proprietary REASONING Anthropic	34.3	—	200	74с	$0.50	Proprietary
37	Grok 4.7 MAX Proprietary REASONING SpaceXAI	34.3	—	84	102с	$6.0	Proprietary
38	Qwen3.8 Max 0902 Proprietary REASONING Alibaba	33.6	—	40	2.9с	$6.0	Proprietary
39	DeepSeek V4.1 Flash 1017 MAX OSS REASONING DeepSeek	32.5	—	170	1.2с	$1.2	Open
40	GLM-5.3 MAX OSS REASONING Z AI	31.7	—	88	1.4с	$4.4	Open
41	Grok 4.7 LOW Proprietary REASONING SpaceXAI	30.9	—	70	6.3с	$6.0	Proprietary
42	Gemini 3.8 Flash HIGH Proprietary REASONING Google	30.5	—	135	25с	$3.8	Proprietary
43	GPT-6.1 Terra Proprietary OpenAI	30.4	—	91	1.0с	$5.0	Proprietary
44	Claude Sonnet 5.5 MED Proprietary REASONING Anthropic	30.3	—	106	1.4с	$10.0	Proprietary
45	Gemini 3.8 Flash MED Proprietary REASONING Google	30.0	—	—	—	$3.8	Proprietary
46	MiMo-V2.6-Pro OSS REASONING Xiaomi	29.9	—	47	3.2с	$0.87	Open
47	Claude Haiku 5.5 HIGH Proprietary REASONING Anthropic	29.6	—	182	32с	$0.50	Proprietary
48	Ling 3.1 Flash 1019 OSS REASONING InclusionAI	29.5	—	—	—	$0.90	Open
49	GPT-6.1 Terra LOW Proprietary REASONING OpenAI	29.2	—	91	1.5с	$5.0	Proprietary
50	MiniMax-M3.1-Flash OSS REASONING MiniMax	28.9	—	136	0.9с	$0.60	Open
51	Step 5 Preview OSS REASONING StepFun	28.5	—	146	0.9с	$2.7	Open
52	DeepSeek V4.1 Flash 1017 HIGH OSS REASONING DeepSeek	28.5	—	—	—	$1.2	Open
53	GLM-5.3-Flash OSS REASONING Z AI	27.8	—	75	1.5с	$0.50	Open
54	GPT-6 Luna MAX Proprietary REASONING OpenAI	27.8	—	118	113с	$0.50	Proprietary
55	MiMo-V2.6-Flash OSS REASONING Xiaomi	26.7	—	38	3.9с	$0.28	Open
56	Mistral Large 4 HIGH OSS REASONING Mistral	26.0	—	—	—	$2.1	Open
57	Claude Sonnet 5.5 LOW Proprietary REASONING Anthropic	25.3	—	107	1.2с	$10.0	Proprietary
58	Muse Spark Open MAX OSS REASONING Meta	24.7	—	196	0.7с	$4.2	Open
59	Claude Haiku 5.5 MED Proprietary REASONING Anthropic	24.6	—	176	14с	$0.50	Proprietary
60	GPT-6 Luna MAX Proprietary REASONING OpenAI	23.3	—	127	40с	$0.50	Proprietary
61	Mistral Large 4 MED OSS REASONING Mistral	23.1	—	—	—	$2.1	Open
62	DeepSeek V4 Pro 0813 MAX OSS REASONING DeepSeek	23.1	—	94	1.8с	$4.0	Open
63	GLM-5.3 LOW OSS REASONING Z AI	23.1	—	82	1.6с	$4.4	Open
64	Gemini 3.8 Flash LOW Proprietary REASONING Google	22.3	—	—	—	$3.8	Proprietary
65	GPT-6 Luna HIGH Proprietary REASONING OpenAI	21.4	—	120	29с	$0.50	Proprietary
66	Qwen3.8 27B MAX OSS REASONING Alibaba	20.1	—	41	4.0с	$3.0	Open
67	MiniMax-M3 OSS REASONING MiniMax	18.7	—	119	0.7с	$1.2	Open
68	Claude Haiku 5.5 LOW Proprietary REASONING Anthropic	18.6	—	168	8.8с	$0.50	Proprietary
69	DeepSeek V4.1 Flash 1017 OSS DeepSeek	18.5	—	185	1.4с	$1.2	Open
70	GPT-6 Luna MED Proprietary REASONING OpenAI	18.3	—	—	—	$0.50	Proprietary
71	K2 Horizon 375B A23B OSS REASONING Institute of Foundation Models	17.8	—	531	2.4с	free	Open
72	Beam 1.0 OSS REASONING Reflection	17.7	—	371	0.6с	$1.2	Open
73	Qwen3.8 27B MED OSS REASONING Alibaba	17.5	—	51	4.1с	$3.0	Open
74	Qwen3.8 27B LOW OSS REASONING Alibaba	15.7	—	48	3.9с	$3.0	Open
75	Nemotron 3.5 Super VL OSS REASONING NVIDIA	13.6	—	485	0.7с	$0.50	Open
76	Inkling Small OSS REASONING Thinking Machines	12.9	—	235	1.6с	$1.2	Open
77	Inkling MAX OSS REASONING Thinking Machines	12.5	—	199	1.7с	$4.0	Open
78	Nemotron 3 Ultra 550B A55B OSS REASONING NVIDIA	11.8	—	161	1.1с	$2.7	Open
79	Gemini 3.5 Flash-Lite Proprietary REASONING Google	11.7	—	423	6.2с	$2.5	Proprietary
80	Qwen3.8 27B OSS Alibaba	10.0	—	46	4.0с	$3.0	Open
81	GPT-6 Luna LOW Proprietary REASONING OpenAI	9.7	—	127	2.8с	$0.50	Proprietary
82	GPT-6 Luna Proprietary OpenAI	9.4	—	118	0.8с	$0.50	Proprietary
83	Muse Glimmer HIGH OSS REASONING Meta	6.3	—	123	1.2с	$1.2	Open
84	Command A+ OSS REASONING Cohere	6.2	—	147	0.5с	free	Open
85	Nemotron 3.5 Lightning OSS REASONING NVIDIA	5.5	—	295	0.6с	$0.20	Open
86	Granite 4.2 3B OSS REASONING IBM	4.7	—	—	—	—	Open
87	Mistral Small 4 OSS REASONING Mistral	4.3	—	167	0.7с	$0.60	Open
88	NVIDIA Nemotron 3 Nano 4B OSS REASONING NVIDIA	1.8	—	—	—	—	Open

Intelligence: Artificial Analysis Blended (MMLU-Pro, GPQA, HLE, IFBench, etc.) Coding: AA Coding Index; «—» означает, что источник не публикует показатель Speed: медианная скорость генерации, токенов/сек TTFT: Time to First Token, секунды (медиана) $/1M out: цена за 1M output-токенов (USD)

Рейтинг Arena AI

Рейтинг на основе слепых попарных сравнений пользователями · 278 open-weight моделей · Данные: официальный датасет Arena AI · CC BY 4.0

#	Модель	Score	CI	Голоса
1	Kimi K3 (Max) moonshot	1488	±4	36K
2	GLM-5.3 (Max) zai	1481	±5	26K
3	MiMo-V2.6-Pro xiaomi	1480	±9	4K
4	GLM-5.3-Flash zai	1477	±5	41K
5	DeepSeek-V4.1-Flash-20260910 (Max) deepseek	1476	±6	13K
6	GLM-5.2 (Max) zai	1475	±4	53K
7	MiMo-V2.5-Pro xiaomi	1467	±4	77K
8	GLM-5.1 zai	1465	±4	66K
9	DeepSeek-V4-Pro (High, 20260813) deepseek	1465	±7	9K
10	Kimi K2.6 moonshot	1461	±4	39K
11	DeepSeek-V4-Pro deepseek	1458	±4	57K
12	GLM-5 zai	1457	±4	29K
13	DeepSeek-V4-Pro Preview (High) deepseek	1455	±4	54K
14	Hy3 tencent	1454	±6	13K
15	Gemma 4 31B google	1452	±7	6K
16	Kimi K2.5 (Thinking) moonshot	1450	±3	74K
17	MiMo-V2.6-Flash xiaomi	1448	±6	15K
18	Inkling thinky	1442	±4	45K
19	Qwen3.5-397B-A17B alibaba	1441	±3	97K
20	GLM-4.7 zai	1441	±6	12K
21	MiniMax-M3 minimax	1439	±4	68K
22	DeepSeek-V4-Flash Preview (High) deepseek	1439	±4	51K
23	Gemma 4 26B A4B google	1438	±7	6K
24	DeepSeek-V4-Flash deepseek	1436	±4	51K
25	Beam 1.0 501B A23B reflection	1436	±9	4K
26	Qwen3.8-27B alibaba	1436	±5	27K
27	MiMo-V2.5 xiaomi	1434	±4	47K
28	Kimi K2.5 (Instant) moonshot	1431	±7	8K
29	Kimi K2 Thinking Turbo moonshot	1430	±3	63K
30	NVIDIA-Nemotron-3-Ultra-550B-A55B (NVFP4) nvidia	1427	±7	11K

227 open-weight моделей · Данные на 28 окт 2026

#	Модель	Score	CI	Голоса
1	Kimi K3 (Max) moonshot	1540	±7	9K
2	MiMo-V2.6-Pro xiaomi	1538	±18	1K
3	GLM-5.3-Flash zai	1526	±7	10K
4	DeepSeek-V4.1-Flash-20260910 (Max) deepseek	1523	±10	3K
5	MiMo-V2.5-Pro xiaomi	1521	±5	20K
6	GLM-5.3 (Max) zai	1520	±8	6K
7	Kimi K2.6 moonshot	1516	±7	11K
8	GLM-5.2 (Max) zai	1514	±6	14K
9	GLM-5.1 zai	1513	±5	18K
10	MiMo-V2.6-Flash xiaomi	1510	±10	3K
11	Beam 1.0 501B A23B reflection	1505	±18	1K
12	DeepSeek-V4-Pro (High, 20260813) deepseek	1505	±11	2K
13	Kimi K2.5 (Instant) moonshot	1504	±14	1K
14	Hy3 tencent	1503	±10	3K
15	Gemma 4 31B google	1502	±15	1K
16	Kimi K2.5 (Thinking) moonshot	1502	±5	20K
17	DeepSeek-V4-Pro deepseek	1500	±6	17K
18	Qwen3.8-27B alibaba	1498	±8	7K
19	GLM-5 zai	1498	±7	7K
20	MiniMax-M3 minimax	1495	±6	19K
21	Inkling thinky	1494	±6	12K
22	Qwen3.5-397B-A17B alibaba	1491	±5	27K
23	MiMo-V2.5 xiaomi	1491	±7	13K
24	DeepSeek-V4-Pro Preview (High) deepseek	1488	±6	16K
25	Kimi K2 Thinking Turbo moonshot	1486	±5	15K
26	GLM-4.7 zai	1484	±12	2K
27	DeepSeek-V4-Flash deepseek	1484	±6	15K
28	Gemma 4 26B A4B google	1482	±15	1K
29	MiniMax-M2.7 minimax	1480	±5	25K
30	DeepSeek-V4-Flash Preview (High) deepseek	1480	±6	15K

222 open-weight моделей · Данные на 28 окт 2026

#	Модель	Score	CI	Голоса
1	GLM-5.3-Flash zai	1271	±6	16K
2	Kimi K2.6 moonshot	1265	±7	16K
3	Gemma 4 31B google	1262	±6	40K
4	DeepSeek-V4.1-Flash-20260910 (Max) deepseek	1261	±10	4K
5	Kimi K2.5 (Thinking) moonshot	1252	±6	28K
6	Qwen3.5-397B-A17B alibaba	1246	±5	32K
7	MiMo-V2.6-Pro xiaomi	1245	±15	1K
8	MiMo-V2.6-Flash xiaomi	1245	±11	3K
9	Qwen3.8-27B alibaba	1242	±8	7K
10	Gemma 4 26B A4B google	1242	±6	23K
11	Kimi K2.5 (Instant) moonshot	1236	±11	4K
12	MiMo-V2.5 xiaomi	1234	±6	29K
13	MiniMax-M3 minimax	1233	±6	19K
14	Qwen3.5-122B-A10B alibaba	1226	±7	13K
15	Qwen3.5-27B alibaba	1217	±6	30K
16	Qwen3-VL-235B-A22B-Instruct alibaba	1215	±7	12K
17	Inkling-Small thinky	1209	±8	8K
18	NVIDIA-Nemotron-3.5-Super-VL-120B-A12B (NVFP4) nvidia	1204	±13	2K
19	Mistral Medium 3.5 mistral	1200	±10	5K
20	Mistral Large 3 mistral	1199	±8	9K
21	Qwen3-VL-235B-A22B-Thinking alibaba	1189	±12	2K
22	GLM-4.6V zai	1161	±14	2K
23	Gemma 3 27B IT google	1160	±8	16K
24	GLM-4.5V zai	1153	±12	3K
25	Llama 4 Maverick 17B 128E Instruct meta	1145	±9	6K
26	Step-3 stepfun	1145	±12	3K
27	Mistral Small 3.2 (2506) mistral	1140	±9	10K
28	Llama 4 Scout 17B 16E Instruct meta	1128	±10	6K
29	Mistral Small 3.1 24B Instruct (2503) mistral	1127	±9	28K
30	Qwen2.5-VL-72B-Instruct alibaba	1121	±11	3K

50 open-weight моделей · Данные на 19 окт 2026

#	Модель	Score	CI	Голоса
1	Qwen-Image-2.1 alibaba	1225	±6	19K
2	Ideogram 4.0 (Quality) ideogram	1200	±4	72K
3	Cosmos3-Super-Text2Image (Agentic) nvidia	1164	±7	10K
4	HunyuanImage 3.0 tencent	1151	±3	180K
5	Cosmos3-Super-Text2Image nvidia	1151	±6	18K
6	FLUX.2 [dev] bfl	1145	±4	77K
7	Qwen-Image (2512) alibaba	1125	±3	98K
8	HiDream-O1-Image hidream	1115	±4	77K
9	Krea 2 Turbo krea	1114	±4	63K
10	Z-Image-Turbo alibaba	1082	±5	24K
11	FLUX.2 [klein] 9B bfl	1070	±3	150K
12	Qwen-Image (Prompt Extend) alibaba	1061	±3	716K
13	Qwen-Image alibaba	1057	±3	84K
14	FLUX.2 [klein] 4B bfl	1030	±3	152K
15	GLM-Image zai	1012	±9	4K
16	FLUX.1 [dev] (FP8) bfl	969	±4	49K
17	FLUX.1 Kontext [dev] bfl	942	±4	216K
18	Stable Diffusion 3.5 Large	938	±4	23K
19	BAGEL bytedance	897	±6	12K

19 open-weight моделей · Данные на 26 окт 2026

#	Модель	Score	CI	Голоса
1	Qwen-Image-2.1 alibaba	1362	±6	39K
2	HunyuanImage 3.0 Instruct tencent	1302	±3	331K
3	Qwen-Image-Edit alibaba	1241	±3	2.0M
4	Qwen-Image-Edit (2511) alibaba	1233	±3	485K
5	FLUX.2 [klein] 9B bfl	1224	±3	540K
6	FLUX.2 [dev] bfl	1224	±4	254K
7	FLUX.2 [klein] 4B bfl	1187	±3	541K
8	FLUX.1 Kontext [dev] bfl	1151	±3	3.6M
9	BAGEL bytedance	1027	±6	13K
10	Step1X-Edit stepfun	996	±5	155K

10 open-weight моделей · Данные на 27 окт 2026

#	Модель	Score	CI	Голоса
1	minimax-h3 minimax	1460	±9	11K
2	kandinsky-5.0-t2v-pro kandinsky	1175	±21	2K
3	hunyuan-video-1.5 tencent	1169	±16	4K
4	ltx-2-19b	1154	±8	75K
5	wan-v2.2-a14b alibaba	1132	±15	10K
6	kandinsky-5.0-t2v-lite kandinsky	1113	±18	1K
7	mochi-v1 genmo	1007	±17	5K

7 open-weight моделей · Данные на 22 сен 2026

#	Модель	Score	CI	Голоса
1	diffbot-small-xl diffbot	1023	±9	6K

1 open-weight моделей · Данные на 13 окт 2026

Свежие релизы

по дате выхода

Mistral Large 4 HIGH

Mistral

OSS

Intel 26.0 29 окт 2026 · $2.1/1M

Gemini 4 Carbon HIGH

Google

Proprietary

Intel 59.4 28 окт 2026 · 95 t/s · $20.0/1M

GPT-6.1 Terra MAX

OpenAI

Proprietary

Intel 51.0 26 окт 2026 · 107 t/s · $5.0/1M

Kimi K3.1 Pro

Kimi

Proprietary

Intel 46.0 23 окт 2026 · 31 t/s · $10.0/1M

MiniMax-M3.1-Flash

MiniMax

OSS

Intel 28.9 19 окт 2026 · 136 t/s · $0.60/1M

DeepSeek V4.1 Flash 1017 MAX

DeepSeek

OSS

Intel 32.5 17 окт 2026 · 170 t/s · $1.2/1M

Beam 1.0

Reflection

OSS

Intel 17.7 16 окт 2026 · 371 t/s · $1.2/1M

Muse Spark Open MAX

Лучшие OSS

Intelligence Index

DeepSeek V4.1 Flash 1017 MAX

DeepSeek

OSS

Intel 32.5 17 окт 2026 · 170 t/s · $1.2/1M

GLM-5.3 MAX

Z AI

OSS

Intel 31.7 18 авг 2026 · 88 t/s · $4.4/1M

MiMo-V2.6-Pro

Xiaomi

OSS

Intel 29.9 21 сен 2026 · 47 t/s · $0.87/1M

Ling 3.1 Flash 1019

InclusionAI

OSS

Intel 29.5 $0.90/1M

MiniMax-M3.1-Flash

MiniMax

OSS

Intel 28.9 19 окт 2026 · 136 t/s · $0.60/1M

Step 5 Preview

StepFun

OSS

Intel 28.5 146 t/s · $2.7/1M

DeepSeek V4.1 Flash 1017 HIGH

DeepSeek

OSS

Intel 28.5 $1.2/1M

GLM-5.3-Flash

Z AI

OSS

Intel 27.8 26 авг 2026 · 75 t/s · $0.50/1M

О проекте

LLM Bench — независимый open-source проект, отслеживающий экосистему языковых моделей. Основной рейтинг строится по Intelligence Index от Artificial Analysis. MMLU-Pro дополняет его академическим срезом, а Arena AI показывает отдельный рейтинг слепых попарных сравнений пользователями.

Данные агрегируются из открытых источников: Artificial Analysis, Arena AI, TIGER-Lab MMLU-Pro и Hugging Face. Обновление происходит автоматически ежедневно. Все результаты проверяемы — ссылки ведут на оригинальные источники.

Проект поддерживается сообществом и не аффилирован ни с одним разработчиком моделей. Исходный код, данные и методология полностью открыты — вклад приветствуется через GitHub.

Методология

Artificial Analysis (основной): Intelligence Index объединяет несколько тестов рассуждения, знаний, инструкций и программирования; также публикуются скорость и цена.
MMLU-Pro: 12 032 вопроса повышенной сложности из 14 предметных областей. Расширенная версия MMLU с 10 вариантами ответов. Дополнительная метрика. Данные: TIGER-Lab.
Агрегация: Artificial Analysis задаёт основной порядок. MMLU-Pro подключается при совпадении модели, Arena AI показывается отдельным блоком. По умолчанию видны только open-weight модели.