Hacker News
Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness
AIFirst seen 158 points55 comments
A open-source project called GVS5H says five copies of one AI model, working together, can match Claude Fable 5 on hard coding tests. The claim is that a smaller open model can reach top-level accuracy when it organizes its own work.
What happened
The GVS5H project shows a method called ledger-based zero-shot self-orchestration. Several fresh copies of one model split a coding problem into parts. They share a folder on a computer that holds a plan, notes, and the current solution. The team tested this on the 100 newest hard problems from LiveCodeBench, a coding test. They say a locally run Qwen3.8-27B model rose from 66.8% to 92.4%.
Key numbers
- Qwen3.8-27B: 66.8% without the method, 92.4% with it.
- Orchestrated Qwen3.8 Flash Next: 93.0%, versus 90.4% for Fable 5.
- The Flash Next run costs about 9% of Fable 5's cost.
- Gains reach up to 25.6 percentage points on pinned backends.
- Eleven open and closed models were tested. Some models did not improve or got worse.
Background
Coding tests like LiveCodeBench give an AI a programming problem and check whether its code works. Usually, the best scores come from very large, expensive models. This project tries a different route. It does not train a new model. Instead, it changes how the same model is used, by splitting the work and keeping shared notes.
Why it matters
If the results hold up, a smaller model you can run yourself might match a costly top model. That could make strong coding help cheaper. The authors also say the gains are not the same for every model, so the method does not work everywhere. Results are from the authors' own tests and have not been checked by others.
What to watch
- Whether other researchers can repeat the 92.4% and 93.0% results.
- Which models gain the most, and which gain nothing.
- How the method performs on tests beyond LiveCodeBench.
Glossary
- LiveCodeBench
- A test that gives AI models new programming problems and checks whether their code works.
- Qwen3.8-27B
- An open AI model with 27 billion parameters that people can run on their own computers.
- Fable 5
- The Claude Fable 5 model, a closed commercial AI used as the top score to beat.
- zero-shot self-orchestration
- A method where one AI model splits up a task and coordinates its own helpers without extra training.