# PostTrainBench v1.2

> Version 1.2 of PostTrainBench, which asks an agent to post-train a base model and scores the resulting model.

Canonical page: https://benchlm.ai/benchmarks/posttrainbench12

- Category: [Coding](/coding)
- Last updated: October 29, 2026

## About PostTrainBench v1.2

- Year: 2026
- Tasks: Agent post-training tasks
- Format: Post-training score
- Difficulty: Autonomous ML engineering
- Paper: [PostTrainBench](https://posttrainbench.com/)

Moonshot runs Kimi K3.1 Pro and DeepSeek-V4.1-Flash-1017 on v1.2 with Claude Code on H20 GPUs instead of the official H100s, judged by DeepSeek-V4.1-Flash-1017 at max effort instead of the official GPT-5.6-Terra judge. BenchLM keeps it separate from the v1.1 key and the unversioned PostTrain Bench key.

PostTrainBench v1.2 is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (1 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Kimi K3.1 Pro](/models/kimi-3-1-pro) | Moonshot AI | 36.5% |

## FAQ

### What does PostTrainBench v1.2 measure?

Version 1.2 of PostTrainBench, which asks an agent to post-train a base model and scores the resulting model.

### Which model scores highest on PostTrainBench v1.2?

Kimi K3.1 Pro by Moonshot AI currently leads with a score of 36.5% on PostTrainBench v1.2.

### How many models are evaluated on PostTrainBench v1.2?

1 AI models have published results on PostTrainBench v1.2 in the BenchLM catalog.

### Does PostTrainBench v1.2 affect BenchLM's overall score?

Not directly. PostTrainBench v1.2 is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.