glm-5.3
GLM route with 100.00 overall score. Best for Data extraction, Instruction following, Math reasoning.
Chinese LLM benchmark
DeepSeek, Qwen, Kimi, GLM, MiniMax and MiMo compared
This English summary turns GPTokens BenchLocal results into a practical shortlist for API buyers. Use it as a screening signal before running your own prompts.
First-tier routes
Top benchmark signals
hy3
Hunyuan route with 99.89 overall score. Best for Instruction following, Math reasoning, Tool calling.
hy4-preview
Hunyuan route with 99.89 overall score. Best for Instruction following, Math reasoning, Tool calling.
deepseek-flash
DeepSeek route with 99.85 overall score. Best for Instruction following, Math reasoning, Tool calling.
GLM 5.2
GLM route with 99.58 overall score. Best for Instruction following, Math reasoning, Tool calling.
BenchLocal V3.2 · four-category mean
Full benchmark table
Sort by
| Model | Total /100 | Extraction | Instructions | Math | Tools | Fully correct | Tokens | Cost (USD) | CPS ($/100 points) |
|---|---|---|---|---|---|---|---|---|---|
glm-5.3glm-5.3 | 100.00 | 100.00 | 100.00 | 100.00 | 100.00 | 60/60 | 112,333 | $0.169500 | $0.169500 |
hy3hy3 | 99.89 | 99.56 | 100.00 | 100.00 | 100.00 | 59/60 | 134,645 | $0.033956 | $0.033994 |
hy4-previewhy4-preview | 99.89 | 99.56 | 100.00 | 100.00 | 100.00 | 59/60 | 224,023 | $0.320884 | $0.321241 |
deepseek-flashdeepseek-flash | 99.85 | 99.39 | 100.00 | 100.00 | 100.00 | 58/60 | 124,027 | $0.051406 | $0.051485 |
GLM 5.2glm-5.2 | 99.58 | 98.33 | 100.00 | 100.00 | 100.00 | 59/60 | 104,610 | $0.165644 | $0.166337 |
deepseek-v4-flash-vision-expdeepseek-v4-flash-vision-exp | 99.54 | 98.17 | 100.00 | 100.00 | 100.00 | 59/60 | 117,481 | $0.047870 | $0.048090 |
kimi-k3kimi-k3 | 99.54 | 98.17 | 100.00 | 100.00 | 100.00 | 59/60 | 101,369 | $0.520768 | $0.523166 |
muse-spark-1.3-contributormuse-spark-1.3-contributor | 99.54 | 98.17 | 100.00 | 100.00 | 100.00 | 59/60 | 166,481 | $0.021288 | $0.021386 |
deepseek-v4.1-flashdeepseek-v4.1-flash | 99.43 | 97.72 | 100.00 | 100.00 | 100.00 | 58/60 | 120,245 | $0.044390 | $0.044644 |
MiMo v2.5mimo-v2.5 | 99.08 | 98.33 | 100.00 | 100.00 | 98.00 | 58/60 | 139,879 | $0.017690 | $0.017854 |
muse-spark-1.2-contributormuse-spark-1.2-contributor | 99.04 | 98.17 | 100.00 | 100.00 | 98.00 | 58/60 | 173,238 | $0.017496 | $0.017665 |
DeepSeek v4 Prodeepseek-v4-pro | 99.00 | 98.33 | 98.67 | 99.00 | 100.00 | 57/60 | 116,757 | $0.145106 | $0.146572 |
glm-5.3-flashglm-5.3-flash | 99.00 | 98.33 | 98.33 | 100.00 | 99.33 | 57/60 | 91,955 | $0.017806 | $0.017986 |
longcat-2.5-preview-freelongcat-2.5-preview-free | 98.93 | 97.72 | 100.00 | 98.00 | 100.00 | 56/60 | 120,513 | $0 | $0 |
qwen3.8-maxqwen3.8-max | 98.83 | 100.00 | 100.00 | 100.00 | 95.33 | 59/60 | 89,876 | $0.255058 | $0.258069 |
space-bunnyspace-bunny | 98.79 | 99.83 | 97.33 | 100.00 | 98.00 | 56/60 | 115,251 | $0.020420 | $0.020670 |
MiMo v2.5 Promimo-v2.5-pro | 98.57 | 96.93 | 100.00 | 100.00 | 97.33 | 52/60 | 140,480 | $0.048874 | $0.049585 |
kimi-k2.7-codekimi-k2.7-code | 98.50 | 100.00 | 100.00 | 100.00 | 94.00 | 59/60 | 85,086 | $0.127630 | $0.129574 |
mimo-v2.6-flashmimo-v2.6-flash | 98.33 | 100.00 | 100.00 | 95.33 | 98.00 | 58/60 | 93,870 | $0.009242 | $0.009399 |
space-bunny-freespace-bunny-free | 98.19 | 98.41 | 98.33 | 98.00 | 98.00 | 53/60 | 114,003 | $0 | $0 |
gpt-6-lunagpt-6-luna | 98.06 | 97.89 | 100.00 | 94.33 | 100.00 | 56/60 | 56,744 | $0.011290 | $0.011514 |
longcat-2.0longcat-2.0 | 97.88 | 91.50 | 100.00 | 100.00 | 100.00 | 58/60 | 121,868 | $0.057866 | $0.059122 |
qwen3.8-flashqwen3.8-flash | 97.72 | 99.56 | 100.00 | 100.00 | 91.33 | 57/60 | 134,617 | $0.028436 | $0.029099 |
gpt-5.6-lunagpt-5.6-luna | 97.42 | 98.33 | 98.67 | 100.00 | 92.67 | 55/60 | 56,458 | $0.025712 | $0.026394 |
grok-4.6grok-4.6 | 96.67 | 100.00 | 100.00 | 100.00 | 86.67 | 58/60 | 171,579 | $0.740338 | $0.765867 |
grok-4.7grok-4.7 | 96.67 | 100.00 | 100.00 | 100.00 | 86.67 | 58/60 | 261,418 | $0.772028 | $0.798650 |
mimo-v2.6-promimo-v2.6-pro | 96.64 | 97.56 | 97.00 | 100.00 | 92.00 | 54/60 | 96,500 | $0.033242 | $0.034398 |
MiniMax M2.7minimax-m2.7 | 95.38 | 97.53 | 100.00 | 84.00 | 100.00 | 55/60 | 88,510 | $0.062790 | $0.065829 |
Qwen 3.7 Plusqwen3.7-plus | 92.72 | 99.56 | 100.00 | 72.00 | 99.33 | 52/60 | 122,334 | $0.116596 | $0.125748 |
MiniMax M3minimax-m3 | 85.88 | 92.96 | 86.89 | 76.33 | 87.33 | 43/60 | 32,257 | $0.010676 | $0.012432 |
BenchLocal V3.2 · Original 27 models: 2 October 2026; three full-suite supplements: 15 October 2026 · 60 tasks per model. Total = mean of the four category scores. The original models retain 45 V3 tasks and 15 revised tool tasks; the three additional model IDs each completed all 60 identical tasks. Near-ceiling scores do not establish general capability; V3 scores are not directly comparable. CPS = Cost per Score; shown as USD per 100 score points: retained execution cost ÷ unrounded total score × 100. Free models are shown at $0. Paid-model costs are historical charges, excluding failed requests and discarded retests.
Production caveat
How to interpret the ranking
V3.2 covers 27 models × 60 tasks, with 15 tasks each for extraction, instruction following, math reasoning and tool calling. The total is their equal-weight mean. Each task retains the first valid execution of its final version; valid wrong answers were not retried for a better score.
All 1,620 retained responses passed the final integrity audit. Each model reused 45 V3 tasks and reran 15 revised tool tasks. Near-ceiling scores are not a reliable general capability ranking; V3 and V3.2 scores are not directly comparable. Tool actions were simulated, and this fixed set does not measure complete coding or autonomous-agent capability.
Public benchmarks should shortlist models, not replace production testing. Before scaling, test latency, fallback behavior, refusal patterns, tool-call reliability, and cost per completed task.
Strong first tests
- GLM 5.3: 100.00 total; 60/60 fully correct.
- HY3: 99.89 total; CPS $0.033994 per 100 points.
- Kimi K3: 99.54 total; read category scores and recorded cost.
- Read the category scores and CPS together, then run your own task set.
FAQ
Questions developers ask before buying Chinese LLM tokens
Is GPTokens run by a model provider?
No. GPTokens is an independent OpenAI-compatible gateway. It issues GPTokens API keys and routes requests to supported Chinese LLM providers through one account.
Can I use an OpenAI SDK?
Yes. Most integrations only need the GPTokens base URL, a GPTokens API key, and a supported model name from the live model list.
Who is this for?
GPTokens is built for developers and teams in the US, Europe, and other global markets that want to test Chinese LLM APIs without managing separate provider accounts, payment methods, and dashboards.