Home / Benchmark Rankings

Benchmark Rankings

Leaderboard system breakdown + quick look at public benchmark performance. All data is time-stamped; scores are subject to the latest edition on each leaderboard's official site.

Three Major Leaderboard Systems

  • SuperCLUE(Monthly): Comprehensive evaluation of Chinese general-purpose LLMs. Currently includes only domestic models. The total score is composed of six dimensions: math reasoning, hallucination control, scientific reasoning, precise instruction following, agent coding, and agent task planning, based on a closed-book private question bank.
  • OpenCompass Compass(Shanghai AI Laboratory): CompassRank composite leaderboard + CompassArena human-preference battle leaderboard, with an open-source, reproducible evaluation framework. Compass 2.0 was released in 2025.
  • International Benchmarks: AIME (math competitions), GPQA-Diamond (science Q&A), SWE-bench Verified (real-world software engineering), etc., used for cross-country comparison between Chinese and overseas models.
Reading Tips: Leaderboards update frequently with differing methodologies. This page provides a snapshot of the tier landscape and public scores; for formal model selection, rely on the latest official reports from each leaderboard plus testing on your own business datasets.

Competitive Tiers of Chinese LLMs

Landscape assessment based on public leaderboards and official technical reports (2026-10)

T1Tier 1
DeepSeek V4.1-Flash / V4-ProKimi K3.1 ProGLM-5.3Qwen3.7-MaxDoubao Seed 2.1 Pro

Since mid-2025, the DeepSeek-R1/V3 series, GLM-4.5, Qwen3, and Kimi-K2 have long held top positions on the SuperCLUE domestic leaderboard; in 2026, V4 / K3 / GLM-5 / Qwen3.7 took over, with the tier structure continuing.

T2Tier 2
MiniMax M3 SeriesTencent Hunyuan Hy4 / Hy3ERNIESpark X1Step-3LongCat-Flash

Possess single-dimension capabilities that can rival Tier 1 in specific areas (long context, Chinese speech, multimodality, cost).

T3Specialized Tier
MiniCPM 4.0InternLM3SenseNova V6Tiangong T1Ling Ling-1T

Differentiated routes such as on-device, academic, and search-augmented models.

Public Benchmark Snapshot

Per official technical reports, public data from 2025–2026, for cross-model reference

BenchmarkWhat It TestsRepresentative Chinese ResultsOverseas Reference
AIME 2025Math CompetitionDeepSeek-R1-0528 ≈87.5; GLM-4.5 98.6 per official figures (with tools); Kimi K2 ≈87.1o3 high ≈96.7
GPQA-DiamondScience Q&AGLM-4.5 79.1；DeepSeek-R1 71.5；Kimi-K2 75.7Grok-4 87.5 / o3 83.3
SWE-bench VerifiedReal-World Software EngineeringKimi-K2 65.8; GLM-4.5 64.2; DeepSeek V4 GA 80.6 (official figures, 2026-07)Claude series leads at around 80
Terminal-Bench 4.0Agent Terminal TasksRefreshed alongside the release of DeepSeek-V4.1-Flash-1017 (2026-10-17)—

Scores are per vendor technical reports or public coverage; testing conditions (e.g., whether tool use is allowed) differ, so results are not directly comparable across models. Overseas model scores are for reference only.

China vs. Global Flagship Comparison

Catalog prices from primary sources + public benchmarks, with all methodologies noted

① Pricing Comparison of Current Flagship APIs

CampModelsInput ($/M)Output ($/M)Output price (Sonnet 5.5 = 1)
Domestic (China)DeepSeek-V4.1-Flash0.11580.46330.06×
Domestic (China)Doubao Seed 2.1 Pro0.38611.93050.25×
Domestic (China)GLM-5.31.08113.39770.44×
Domestic (China)Qwen3.7-Max1.27413.82320.50×
Domestic (China)Kimi K3.1 Pro1.54447.72201.00×
InternationalGemini 4 Argon1.54447.72201.00×
InternationalGPT-6.1 Sol1.54447.72201.00×
InternationalClaude Sonnet 5.51.54447.72201.00× (baseline)
InternationalClaude Opus 4.73.861019.30502.50×
InternationalClaude Fable 5.17.722038.61005.00×

Same catalog, same-time snapshot (2026-10, official list prices), consistent methodology for direct comparison. Domestic flagships' output prices are roughly 1/2 to 1/16 of comparable international flagships, with an even wider gap for DeepSeek's budget tier.

② Public Benchmarks: Best Domestic vs International Flagships

BenchmarkBest domestic (public figures)Overseas ReferenceStatus
AIME 2025 mathGLM-4.5 98.6 (with tools); DeepSeek-R1-0528 87.5o3 high 96.7On par with peers
SWE-bench Verified codingDeepSeek V4 GA 80.6 (official figures, 2026-07)Claude family ≈80Roughly even
GPQA-Diamond scienceGLM-4.5 79.1；Kimi-K2 75.7Grok-4 87.5；o3 83.3Behind by about 8 points
Overall leaderboard rankingDomestic first tier near the topClaude Opus 5.5 / Gemini 4 Argon / Claude Fable 5.1 hold the top three (per aggregator sites, 2026-10)International models still lead

③ How to Read This Comparison

  • Coding and math: the generational gap has been closed.In 2025, Chinese flagships lagged international models by a generation; in 2026, DeepSeek V4 scores 80.6 on SWE-bench per official figures, with public reports calling it "close to Opus 4.8," and math benchmarks on par with o3.
  • Price: the clearest competitive advantage.In the same catalog, domestic output prices are about 1/2–1/16 of comparable international models; third-party reviews put V4 pricing at roughly 1/57 of Fable 5 (2026-07 figures).
  • Gaps remain:Scientific reasoning (GPQA) trails by about 8 points; the top three on overall leaderboards are still international models; and international players remain the default for toolchain ecosystems, multilingual support, and enterprise services.
  • Selection takeaway:For Chinese-language scenarios and cost-sensitive workloads, domestic flagships are the default choice; for cutting-edge capability and global compliance, international flagships still lead.

Benchmark scores come from vendors' technical reports and public coverage; testing conditions differ, so they cannot be compared directly across sources. International scores are for reference only. Prices may change at any time—refer to official pages for current pricing.

Open-Source Tier

The landscape of domestic models with publicly downloadable weights

Open-Source T1Comprehensively ahead
DeepSeek V4 series (MIT)Kimi K3（2.8T）Qwen3 Open-Source FamilyGLM Open-Source Line

A regular at the top of HuggingFace Trending upon release; Qwen-derived models lead the world in both count and downloads (2025 official figures: 100K+ derived models).

Open-Source T2Each Has Its Strengths
Hunyuan Hy4 PreviewLongCat-FlashLing Ling-1TStep-3ERNIE 4.5InternLM3

Internet giants and academic labs are open-sourcing intensively, covering MoE, multimodal, and lightweight on-device models in multiple sizes.

The official HuggingFace Open LLM Leaderboard stopped updating in 2025; current open-source buzz is best gauged by Trending, download counts, and LMArena.