Qwen3.8-27B edges past Fable 5 by passing itself notes

Researchers lifted Qwen3.8-27B, an open-weight model served locally, from 69.2% to 92.4% pass@1 on the 100 latest hard LiveCodeBench problems without retraining it, according to a paper and code published on GitHub on September 12. That puts the 27B model slightly ahead of Claude Fable 5, which scores 90.4% on the same set.

The authors call the method ledger-based zero-shot self-orchestration. Fresh instances of a single model break a problem into parts and coordinate through a shared filesystem. That filesystem holds a plan, notes and the current solution. No instance has to carry the whole job in its own context: each one reads the ledger, does its piece and writes back. The paper’s transcript analysis attributes the gain to decomposition and persistent context.

The comparison covers nine open and closed-weight models. On pinned backends, the paper reports gains of up to 23.2 percentage points, and Qwen3.8-27B’s own climb works out to the same 23.2 points. The authors describe its result against Fable 5 as “slightly exceeding” it.

The second route goes through the cloud and is about price. Orchestrated GPT-5.6-Terra reaches 88.0% pass@1, still short of Fable 5’s 90.4%, at 19% of Fable 5’s cost.

Not every model gains, and the authors say so: some of the nine came out unchanged under orchestration, and some got worse. The same scaffold that lifts Qwen3.8-27B by more than 20 points can leave another model flat or behind its single-call baseline.

LiveCodeBench keeps adding new problems from contests on LeetCode, AtCoder and Codeforces, and it tags each problem with a release date, so a model can be tested on problems that came out after its training cutoff. The run command in the repository sets LCB_RELEASE=release_v6, which the LiveCodeBench repository lists as 1,055 problems collected through April 2025.

A similar effect showed up on a different benchmark earlier in September, when a harness swap took GPT-6 Astra from 62.7% to 99.9% on ARC-AGI-3.

Reproducing a run takes uv and an API key for the model under test. From codebase/v2-current, the script escalation/run_bench.py runs with --engine multiagent for the orchestrating manager or --engine single for the one-call baseline. The authors’ example uses MULTIAGENT_MODEL=openai:gpt-5.6-terra; other providers plug in as anthropic:<model>, dashscope:<model> or openrouter:<model>, each with its own *_API_KEY.

The pass@1 score prints at the end of a run. Results go to runs/results.json, and the workspaces, where the plans, notes and solutions the instances passed to each other end up, go to runs/ws/. The code is under the MIT License. The paper, figures and run data are under CC BY 4.0.

“Inference-time organization can approach frontier coding accuracy at a fraction of the cost, or slightly exceed it on self-hostable weights,” the authors write.