Coding benchmark

Terminal-Bench leaderboard

13 models have a published score. DeepSeek V4.1 Flash (DeepSeek) leads with 90.6, and the spread to the lowest score, Nemotron 3 Ultra at 56.4, is 34.2 points. The highest-scoring model you can deploy on OpenLLM Buddy is DeepSeek V4.1 Flash, ranked 1 of 13 with 90.6. Sources report different settings for this test, such as test year, version, or reasoning effort. Each setting is listed in the source table, so compare close scores with care.

Best openllm buddy catalog

90.6

DeepSeek V4.1 Flash

6 models scored

Best open weights

88.2

GLM 5.3

7 models scored

Best closed api

No published score

0 models scored

Terminal-Bench scores

Percentage of tasks completed. The benchmark version is shown where the source states it.

  • OpenLLM Buddy catalog
  • Open weights
  • Closed API
  1. 1DeepSeek V4.1 FlashTerminal-Bench 2.1, Pass@1, max effort90.6
  2. 2MiMo V2.6 ProTerminal Bench 2.189.9
  3. 3Kimi K3Terminal-Bench 2.1, max effort88.3
  4. 4GLM 5.3Terminal-Bench 2.188.2
  5. 5DeepSeek V4 ProTerminal-Bench 2.1, V4-Pro-0813, max effort, DeepSeek Harness87.9
  6. 6Tencent Hy4 PreviewTerminal-Bench 2.185.4
  7. 7GLM 5.3 FlashTerminal-Bench 2.1 (Z.ai vendor-reported)84.3
  8. 8Qwen3.8 2.4T-A95BTerminal-Bench 2.1 (Terminus 2 agent), open checkpoint Qwen3.8-2.4T-A95B (xhigh); competitor comparison column82.0
  9. 9Reflection Beam 1.0Terminal-Bench 2.1, BF1680.1
  10. 10Muse Spark OpenTerminal-Bench 2.1, Terminus 2 agent, xhigh effort76.4
  11. 11Qwen3.8 27BTerminal Bench 2.1 (Terminus harness)73.0
  12. 12Qwen3.6 27BTerminal-Bench 2.059.3
  13. 13Nemotron 3 UltraTerminal-Bench 2.156.4

What it measures

Tasks such as compiling a project, configuring a server, fixing a broken environment, or training a small model, each run in a sandboxed container and checked by scripts.

How it is scored

Percentage of tasks completed. The benchmark version is shown where the source states it.

Why it matters

It reflects how coding agents actually work: by running commands and reacting to output.

Official benchmark page

More coding tests

Sources

ModelTestScoreSettingSource
DeepSeek V4.1 FlashTerminal-Bench90.6Terminal-Bench 2.1, Pass@1, max effortDeepSeek-V4.1-Flash model card (Hugging Face)
MiMo V2.6 ProTerminal-Bench89.9Terminal Bench 2.1MiMo-V2.6-Pro-RL model card (Hugging Face)
Kimi K3Terminal-Bench88.3Terminal-Bench 2.1, max effortKimi K3 model card (Hugging Face)
GLM 5.3Terminal-Bench88.2Terminal-Bench 2.1Hugging Face model card (GLM-5.3)
DeepSeek V4 ProTerminal-Bench87.9Terminal-Bench 2.1, V4-Pro-0813, max effort, DeepSeek HarnessHugging Face model card (DeepSeek-V4-Pro-0813)
Tencent Hy4 PreviewTerminal-Bench85.4Terminal-Bench 2.1Hugging Face model card (tencent/Hy4-preview), evaluation results
GLM 5.3 FlashTerminal-Bench84.3Terminal-Bench 2.1 (Z.ai vendor-reported)DataNorth news (quoting Z.ai official results)
Qwen3.8 2.4T-A95BTerminal-Bench82.0Terminal-Bench 2.1 (Terminus 2 agent), open checkpoint Qwen3.8-2.4T-A95B (xhigh); competitor comparison columnMeta Muse Spark Open model card (comparison column)
Reflection Beam 1.0Terminal-Bench80.1Terminal-Bench 2.1, BF16Hugging Face model card (Beam-1.0-501B-A23B)
Muse Spark OpenTerminal-Bench76.4Terminal-Bench 2.1, Terminus 2 agent, xhigh effortHugging Face model card (Muse-Spark-Open-816B-A42B)
Qwen3.8 27BTerminal-Bench73.0Terminal Bench 2.1 (Terminus harness)Qwen3.8-27B model card (Hugging Face)
Qwen3.6 27BTerminal-Bench59.3Terminal-Bench 2.0Qwen3.6-27B model card (Hugging Face)
Nemotron 3 UltraTerminal-Bench56.4Terminal-Bench 2.1Hugging Face model card (NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16)

Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works