Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Coding benchmark
PostTrainBench v1.2. Every model the catalog carries a published PostTrainBench value for, ranked by that value.
An AI agent gets ten hours and one H100 GPU to post-train a small base model; the score is how well the trained models do on six benchmarks. 21 models in the ModelScale catalog have a published value.
Each point is a model with a published PostTrainBench value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
21 models with a published PostTrainBench value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Weighted aggregate across four base models and six benchmarks |
|---|---|---|---|
| 1 | Anthropic | 43.8 | |
| 2 | OpenAI | 41.9 | |
| 3 | 41.0 |
What the benchmark measures and how its value is published, from the source’s own record.
PostTrainBench (PostTrainBench v1.2) is a coding benchmark published by PostTrainBench. An AI agent gets ten hours and one H100 GPU to post-train a small base model; the score is how well the trained models do on six benchmarks. Its evaluation set is described by the source as post-training four base models on six benchmarks. The source rates its difficulty as autonomous ML engineering.
Each model's result is published by PostTrainBench as weighted aggregate across four base models and six benchmarks. ModelScale prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 21 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.
| 4 | 40.4 |
| 5 | OpenAI | 38.8 |
| 6 | Z.AI | 36.8 |
| 7 | Moonshot AI | 36.5 |
| 8 | Anthropic | 35.7 |
| 9 | Tencent | 35.2 |
| 10 | Z.AI | 34.5 |
| 11 | TMInkling 2 Small | Thinking Machines Lab | 31.9 |
| 12 | Moonshot AI | 31.5 |
| 13 | OpenAI | 31.0 |
| 14 | Anthropic | 31.0 |
| 15 | Z.AI | 30.4 |
| 16 | Anthropic | 24.3 |
| 17 | Meta | 24.2 |
| 18 | OpenAI | 23.9 |
| 19 | xAI | 22.6 |
| 20 | 18.1 |
| 21 | OpenAI | 17.8 |
Evidence key: Observed
Rows are ordered by the value PostTrainBench published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.