Kimi K2.7 post-trained on 2,400 coding tasks lifts SWE-Bench Pro to 68.1%
echen · x · 2026-10-22
Kimi team member echen shared results of a new post-training run on Kimi K2.7 using 2,400 coding tasks:
- DeepSWE: 31% → 48.7% (+17.7pp)
- Terminal Bench 4: 0% → 14.3% (+14.3pp)
- SWE-Bench Pro: 60.1% → 68.1% (+8.0pp)
The numbers suggest targeted coding-task post-training yields sizable gains on agentic coding benchmarks, most notably on Terminal Bench 4, which started near zero.
Related event: Kimi K2.7 Post-Training on 2,400 Coding Tasks Boosts Benchmarks(2 posts)→
More from coding & agent
- A voice-driven MCP chain: ChatGPT Voice orchestrating multiple CLI coding agents — PurposePossible1649 · 2026-10-22
- Kornia now requires PRs to disclose which part was done by a human — kornia_foss · 2026-10-22
- Classic 'too many tabs' meme applied to coding agents: send help — mrdrozdov · 2026-10-22
- Open-source sample builds durable audit trails into AI phone workflows via an API — AIBotFromFuture · 2026-10-22
- YC F26 startup Latent reads model internals to catch hallucinations at 530x lower cost — ycombinator · 2026-10-22
- We need an L0.5 for continual learning: agent state stacks up in 4 layers — willcb · 2026-10-22