Skip to content
Weights Not Walls
Menu

Surface Evolver Bench

Reported

No description yet: we only describe an eval set when we can do it correctly.

Kind
Published by a third party's leaderboard, collected by Epoch AI. The leaderboard is named and linked.
Link
https://yhenon.github.io/surface-evolver-llm-eval/
Score shown
Epoch's column "Mean score", as a percentage, higher is better; random guessing scores about 0.0%. The source gives no error bars.
Results
29 in all: 16 for open-weight models, 13 for closed ones. 13 linked to a model page here.

Best open-weight result: moonshotai/Kimi-K3, 95.0%, number 2 of 29 results.

Surface Evolver Bench results
#ModelWeightsSettingScoreDateSource
1claude-fable-5_highAnthropicclosedeffort high95.0%not datedyhenon.github.io
2moonshotai/Kimi-K3Epoch: kimi-k3_unknown, Moonshotopeneffort unknown95.0%not datedyhenon.github.io
3gpt-5.6-sol_xhighOpenAIclosedeffort xhigh93.1%not datedyhenon.github.io
4moonshotai/Kimi-K3Epoch: kimi-k3_max, Moonshotopeneffort max93.0%not datedyhenon.github.io
5gpt-5.5_highOpenAIclosedeffort high88.1%not datedyhenon.github.io
6claude-opus-4-8_highAnthropicclosedeffort high87.5%not datedyhenon.github.io
7gpt-5.6-terra_xhighOpenAIclosedeffort xhigh83.8%not datedyhenon.github.io
8gpt-5.5_mediumOpenAIclosedeffort medium81.3%not datedyhenon.github.io
9gemini-3.8-flash_highGoogle DeepMindclosedeffort high76.9%not datedyhenon.github.io
10grok-4.5_highxAIclosedeffort high74.4%not datedyhenon.github.io
11claude-opus-4-8_noneAnthropicclosedeffort none68.1%not datedyhenon.github.io
12gpt-5.6-luna_mediumOpenAIclosedeffort medium61.9%not datedyhenon.github.io
13claude-sonnet-5_mediumAnthropicclosedeffort medium60.0%not datedyhenon.github.io
14gemini-3.5-flash_mediumGoogle DeepMindclosedeffort medium58.1%not datedyhenon.github.io
15zai-org/GLM-5.2Epoch: glm-5.2_high, Z.ai (Zhipu AI)openeffort high55.6%not datedyhenon.github.io
16MiniMaxAI/MiniMax-M3Epoch: MiniMax-M3, MiniMaxopendefault55.0%not datedyhenon.github.io
17zai-org/GLM-5.3-FlashEpoch: glm-5.3-flash_max, Z.ai (Zhipu AI)openeffort max52.5%not datedyhenon.github.io
18muse-spark-1.1_highMeta AIclosedeffort high52.5%not datedyhenon.github.io
19moonshotai/Kimi-K2.7-CodeEpoch: kimi-k2.7-code, Moonshotopendefault48.8%not datedyhenon.github.io
20deepseek-ai/DeepSeek-V4.1-FlashEpoch: deepseek-v4.1-flash_high, DeepSeekopeneffort high46.3%not datedyhenon.github.io
21Qwen/Qwen3.8-27BEpoch: qwen3.8-27b_xhigh, Alibabaopeneffort xhigh45.0%not datedyhenon.github.io
22Qwen/Qwen3.6-35B-A3BEpoch: qwen3.6-35b-a3b, Alibabaopendefault44.4%not datedyhenon.github.io
23deepseek-ai/DeepSeek-V4-ProEpoch: deepseek-v4-pro_high, DeepSeekopeneffort high40.0%not datedyhenon.github.io
24google/gemma-4-31B-itEpoch: gemma-4-31b-it, Google DeepMindopendefault30.6%not datedyhenon.github.io
25mistral-large-4-0_unknownMistral AIno page matchopeneffort unknown28.1%not datedyhenon.github.io
26mistral-medium-2604Mistral AIno page matchopendefault26.9%not datedyhenon.github.io
27openai/gpt-oss-120bEpoch: gpt-oss-120b, OpenAIopendefault25.0%not datedyhenon.github.io
28poolside/Laguna-M.1Epoch: laguna-m.1, Poolsideopendefault15.6%not datedyhenon.github.io
29trinity-large-thinkingArcee AIno page matchopendefault15.6%not datedyhenon.github.io

Open-weight rows are shaded. Dates are when the result was run or evaluated, where the data says. Data from Epoch AI (CC-BY 4.0), fetched 19 Nov 2026: Epoch AI, ‘Capabilities & benchmarking’. Published online at epoch.ai. Retrieved from ‘https://epoch.ai/benchmarks’ [online resource]. Epoch AI. Method