Open LLM Leaderboard

Name: Open LLM Leaderboard - LLM Stats
Creator: LLM Stats
License: https://llm-stats.com/legal/terms-of-service

Ranking the best open LLMs by performance, price, and speed

Select 2-4 models to compare

← Scroll →


Kimi K3NEW	55.6	🇨🇳	Open	1.0M	$3.00	$15.00	0c/s	–	54.9	42.3	44.8	34.4	–	39.5	39.1	–	32.1	32.1	–	93.5%	—	—	—	—	—	91.2%	91.3%	81.6%	—	84.2%	56.0%	—	—	73.2%	—	—	—	—	—	37.6%	—	2800	23.0s	-	Jul. 2026	MoonshotAI
GLM-5.2	47.6	🇨🇳	Open	1.0M	$0.95	$3.00	14c/s	–	46.7	42.0	39.9	–	–	28.7	26.0	–	–	–	–	91.2%	—	—	—	—	—	—	—	—	—	76.8%	54.7%	—	—	48.2%	—	—	—	—	—	—	62.1%	753	5.7s	-	Jun. 2026	ZAI
Kimi K2.6	44.9	🇨🇳	Open	262.1k	$0.75	$3.50	53c/s	–	45.0	38.0	35.6	26.7	–	30.9	23.5	–	25.0	20.4	–	90.5%	—	80.2%	—	—	—	86.3%	86.7%	80.1%	—	—	36.4%	—	—	50.0%	—	—	—	—	52.2%	27.9%	58.6%	1000	1.5s	-	Apr. 2026	MoonshotAI
Hy3	44.0	🇨🇳	Open	-	-	-	–	–	44.0	35.8	36.2	25.1	–	–	25.9	26.8	–	–	–	90.4%	—	78.0%	—	—	—	84.2%	—	—	—	79.1%	—	—	—	48.5%	—	—	—	—	—	25.6%	57.9%	295	–	-	Jul. 2026	Tencent
DeepSeek-V4-Pro-Max	44.2	🇨🇳	Open	1.0M	$1.60	$3.20	18c/s	–	44.4	41.3	33.9	19.3	–	23.3	25.4	9.7	31.5	31.5	31.3	90.1%	—	80.6%	—	—	—	83.4%	—	—	—	73.6%	48.2%	57.9%	—	51.8%	—	—	—	—	—	—	55.4%	1600	5.5s	-	Apr. 2026	DeepSeek
Qwen3.5-397B-A17B	39.6	🇨🇳	Open	-	-	-	–	–	39.4	37.6	23.1	19.2	17.3	19.8	18.2	28.3	32.1	32.1	38.7	88.4%	—	76.4%	—	88.5%	—	69.0%	—	—	—	—	28.7%	—	—	38.3%	—	—	—	—	—	—	—	397	–	-	Feb. 2026	Qwen
DeepSeek-V4-Flash-Max	39.8	🇨🇳	Open	1.0M	$0.10	$0.20	6c/s	–	40.6	40.5	29.5	13.8	–	21.5	19.5	3.9	28.6	28.6	28.1	88.1%	—	79.0%	—	—	—	73.2%	—	—	—	69.0%	45.1%	34.1%	—	47.8%	—	—	—	—	—	—	52.6%	284	12.7s	-	Apr. 2026	DeepSeek
Qwen3.6-27B	36.7	🇨🇳	Open	262.1k	$0.60	$3.60	221c/s	–	37.2	35.4	27.1	–	–	26.0	15.3	18.8	29.4	29.4	36.1	87.8%	—	77.2%	—	—	82.9%	—	78.4%	75.8%	—	—	24.0%	—	—	—	—	—	—	—	—	—	53.5%	27.8	968ms	-	Apr. 2026	Qwen
Kimi K2.5	39.8	🇨🇳	Open	-	-	-	–	–	39.7	37.2	25.0	22.4	–	29.1	7.5	27.5	30.6	30.6	35.4	87.6%	96.1%	76.8%	—	—	—	74.9%	77.5%	78.5%	—	—	50.2%	—	—	—	—	—	—	—	48.7%	—	50.7%	1000	–	-	Jan. 2026	MoonshotAI
Nemotron 3 Ultra (550B A55B)	37.4	🇺🇸	Open	-	-	-	–	–	36.7	36.9	22.0	3.4	16.3	17.6	0.4	24.6	28.2	30.4	34.0	87.0%	—	70.7%	—	—	—	44.4%	—	—	—	—	37.4%	—	—	—	—	—	—	—	44.6%	—	—	550	–	Sep. 2025	Jun. 2026	Nvidia
Qwen3.5-122B-A10B	35.9	🇨🇳	Open	-	-	-	–	–	36.3	37.0	20.7	16.4	14.5	26.9	12.2	25.9	30.5	30.5	38.0	86.6%	—	72.0%	—	86.7%	83.9%	63.8%	77.2%	76.9%	70.4%	—	47.5%	—	—	—	—	—	—	—	—	—	—	122	–	-	Feb. 2026	Qwen
GLM-5.1	40.9	🇨🇳	Open	200k	$1.40	$4.40	3c/s	–	40.6	34.5	33.7	17.7	–	27.1	21.5	–	25.1	21.5	–	86.2%	—	—	—	—	—	79.3%	—	—	—	71.8%	52.3%	—	—	40.7%	—	—	—	—	—	—	58.4%	754	2.1s	-	Apr. 2026	ZAI
Qwen3.6-35B-A3B	33.5	🇨🇳	Open	-	-	-	–	–	34.2	32.2	21.0	7.3	–	24.4	12.2	20.6	26.9	26.9	33.5	86.0%	—	73.4%	—	—	81.7%	—	78.0%	75.3%	—	62.8%	21.4%	—	—	26.9%	—	—	—	—	—	—	49.5%	35	–	-	Apr. 2026	Qwen
GLM-4.7	35.3	🇨🇳	Open	-	-	-	–	–	35.2	35.1	17.9	11.9	–	20.6	8.0	–	25.3	25.3	25.3	85.7%	95.7%	73.8%	—	—	—	52.0%	—	—	—	—	42.8%	—	—	—	33.3%	—	—	—	—	—	—	358	–	-	Dec. 2025	ZAI
Qwen3.5-27B	34.8	🇨🇳	Open	262.1k	$0.30	$2.40	5c/s	–	35.4	35.8	17.2	15.4	13.6	25.5	12.1	25.2	31.9	31.9	35.3	85.5%	—	72.4%	—	85.9%	82.3%	61.0%	79.5%	75.0%	70.3%	—	48.5%	—	—	—	—	—	—	—	—	—	—	27	12.8s	-	Feb. 2026	Qwen
Kimi K2-Thinking-0905	36.9	🇨🇳	Open	-	-	-	–	–	37.1	38.6	21.7	15.1	-5.3	25.6	–	–	19.1	19.0	31.9	84.5%	100.0%	71.3%	—	—	—	60.2%	—	—	—	—	51.0%	—	—	—	47.1%	—	—	—	44.8%	—	—	1000	–	-	Sep. 2025	MoonshotAI
Gemma 4 31B	33.5	🇺🇸	Open	262.1k	$0.13	$0.38	5c/s	–	33.9	30.7	–	–	–	20.4	13.7	18.3	19.8	19.8	27.3	84.3%	—	—	—	88.4%	—	—	—	76.9%	—	—	26.5%	—	—	—	—	—	—	66.4%	—	—	—	30.7	26.2s	Jan. 2025	Apr. 2026	Google
Qwen3.5-35B-A3B	31.7	🇨🇳	Open	-	-	-	–	–	32.8	33.5	12.1	12.5	11.9	22.9	11.8	21.8	30.1	30.1	32.4	84.2%	—	69.2%	—	85.2%	81.4%	61.0%	77.5%	75.1%	68.6%	—	47.4%	—	—	—	—	—	—	—	—	—	—	35	–	-	Feb. 2026	Qwen
MiMo-V2-Flash	31.6	🇨🇳	Open	-	-	-	–	–	31.7	30.3	17.8	10.2	–	10.4	5.8	15.9	26.4	26.4	26.4	83.7%	94.1%	73.4%	—	—	—	58.3%	—	—	—	—	22.1%	—	—	—	30.5%	—	—	—	—	—	—	309	–	-	Dec. 2025	Xiaomi
Nemotron 3 Super (120B A12B)	27.5	🇺🇸	Open	-	-	-	–	–	25.6	31.3	13.4	-0.9	10.0	10.8	3.2	15.2	26.9	26.9	27.7	82.7%	90.2%	53.7%	—	—	—	31.3%	—	—	—	—	22.8%	—	—	—	25.8%	—	—	—	42.0%	—	—	120	–	Jun. 2025	Mar. 2026	Nvidia
DeepSeek-V3.2 (Thinking)	33.7	🇨🇳	Open	-	-	-	–	–	33.6	31.4	23.4	10.4	–	12.5	10.3	–	26.5	26.5	26.5	82.4%	93.1%	73.1%	—	—	—	51.4%	—	—	—	—	25.1%	—	—	35.2%	—	—	—	—	—	—	—	685	–	-	Dec. 2025	DeepSeek
DeepSeek-V3.2	34.1	🇨🇳	Open	-	-	-	–	–	34.1	32.3	23.5	10.5	–	19.7	11.5	–	26.6	26.6	26.6	82.4%	93.1%	73.1%	—	—	—	51.4%	—	—	—	—	40.8%	—	—	35.2%	—	—	—	—	—	—	—	685	–	-	Dec. 2025	DeepSeek
Gemma 4 26B-A4B	30.3	🇺🇸	Open	262.1k	$0.13	$0.40	33c/s	–	30.2	26.1	–	–	–	15.1	12.6	18.3	22.5	22.5	24.5	82.3%	—	—	—	86.3%	—	—	—	73.8%	—	—	17.2%	—	—	—	—	—	—	44.1%	—	—	—	25.2	2.5s	Jan. 2025	Apr. 2026	Google
Qwen3.5-9B	25.7	🇨🇳	Open	-	-	-	–	–	25.6	24.8	–	–	10.3	12.6	10.4	18.9	23.8	23.8	24.5	81.7%	—	—	—	81.2%	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	9	–	-	Mar. 2026	Qwen
LongCat-Flash-Thinking	29.7	🇨🇳	Open	-	-	-	–	–	29.8	28.8	14.8	–	18.1	8.1	17.1	–	27.6	22.7	22.7	81.5%	90.6%	59.4%	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	560	–	-	Sep. 2025	Meituan
Qwen3-235B-A22B-Thinking-2507	29.0	🇨🇳	Open	-	-	-	–	–	29.4	32.1	–	–	15.7	13.5	11.6	–	32.0	31.9	32.0	81.1%	92.3%	—	—	—	—	—	—	—	—	—	18.2%	—	—	—	—	67.8%	—	—	—	—	—	235	–	-	Jul. 2025	Qwen
DeepSeek-R1-0528	25.6	🇨🇳	Open	-	-	-	–	–	25.2	27.2	9.5	-10.1	–	6.5	–	–	26.3	26.3	26.3	81.0%	87.5%	44.6%	—	—	—	8.9%	—	—	—	—	17.7%	92.3%	—	—	5.7%	—	—	—	—	—	—	671	–	-	May 2025	DeepSeek
GLM-4.6	30.2	🇨🇳	Open	-	-	-	–	–	30.1	25.7	15.5	4.8	–	6.2	–	–	–	–	–	81.0%	93.9%	68.0%	—	—	—	45.1%	—	—	—	—	17.2%	—	—	—	40.5%	—	—	—	—	—	—	357	–	-	Sep. 2025	ZAI
MiniMax M2.1	33.0	🇨🇳	Open	1M	$0.30	$1.20	12c/s	–	33.2	28.1	22.8	11.8	13.7	10.0	16.2	14.1	32.8	32.8	32.8	81.0%	81.0%	67.0%	—	—	—	62.0%	—	—	—	—	22.0%	—	—	43.5%	47.9%	—	—	—	39.0%	—	—	230	2.8s	-	Dec. 2025	MiniMax
GPT OSS 120B High	26.7	🇺🇸	Open	-	-	-	40c/s	–	26.4	24.5	–	–	–	–	2.6	–	18.9	18.9	18.9	80.9%	92.5%	—	—	83.8%	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	116.8	12.1s	-	Aug. 2025	OpenAI

Showing 1–30 of 192 models

Open LLM Leaderboard highlights

Independent ranking of open-weight large language models — Llama, Qwen, GLM, DeepSeek, Mistral, Kimi and more — by coding-arena score, GPQA Diamond, throughput, latency, and per-token pricing. Updated continuously from provider APIs and verified benchmarks. See the LLM Stats Score methodology for how rankings are computed.

Best on GPQA Diamond: GLM-5.2 (91.2%)
Best on AIME 2025: Kimi K2-Thinking-0905 (100.0%)
Best on SWE-Bench Verified: DeepSeek-V4-Pro-Max (80.6%)
Highest throughput: Qwen3.6-27B (221 tok/s)
Lowest latency: GPT OSS 120B (529.80 s)
Cheapest input: Nemotron 3 Nano (30B A3B) ($0.06 / 1M tok)
Largest context window: GLM-5.2 (1.0M tokens)

FAQ

Common questions about the open llm leaderboard

What is the best open-source LLM right now?

Based on coding-arena performance — the most discriminating signal at the frontier — the top model currently leads. For knowledge-heavy reasoning (GPQA Diamond), GLM-5.2 scores highest. Choose by axis rather than a single ranking — see the highlights above for per-metric leaders.

How does the Open LLM Leaderboard rank models?

Models are sorted by coding-arena score (when available), then by GPQA Diamond. Each row aggregates verified benchmark results, provider-reported pricing, and live performance metrics (output throughput and time-to-first-token) sampled across the major API providers. See the LLM Stats Score methodology for the full weighting and refresh cadence.

How many models are tracked?

This leaderboard tracks 191 canonical models across every major lab and provider. New releases typically appear within hours.

Where does pricing data come from?

Per-model input/output pricing is pulled from each provider's public API price list and verified against billing samples from the LLM Stats proxy. When a model is hosted by multiple providers, the cheapest available rate is shown by default.

How is performance measured?

Output throughput (tokens/second) and time-to-first-token are measured by routing standardized prompts through each provider's API and averaging over a 7-day rolling window. Numbers update hourly. Per-model splits live on each model detail page.

How often does the data update?

Pricing and model metadata revalidate every hour. Live performance metrics update on a 7-day rolling average. Benchmark scores update when a new verified result is published or a new evaluation lands on LLM Stats.