{"slug": "gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard", "title": "GVS5H: Five Qwen3.8-27B Models Match Claude Fable 5 on LiveCodeBench Hard", "summary": "A training-free method called ledger-based zero-shot self-orchestration lifted open-weight Qwen3.8-27B from 69.2% to 92.4% pass@1 on the 100 latest hard LiveCodeBench problems, slightly exceeding Claude Fable 5's 90.4%, according to the GVS5H project. Orchestrated GPT-5.6-Terra reached 88.0% pass@1 at 19% of Fable 5's cost, and the method produced gains of up to 23.2 percentage points across nine open and closed-weight models, though some models were unchanged or worse. Transcript analysis attributed the gains to problem decomposition and persistent context, with code released under the MIT License and the paper and run data under CC BY 4.0.", "body_md": "**Abstract.** Frontier coding performance is typically bought with larger proprietary models at high cost. We introduce ledger-based zero-shot self-orchestration, a training-free method in which fresh instances of one model decompose problems and coordinate through a shared filesystem holding a plan, notes and current solution. Across nine open and closed-weight models on the 100 latest *hard* LiveCodeBench problems, the method yields gains of up to 23.2 percentage points on pinned backends and offers two routes to frontier-level accuracy. Orchestrated GPT-5.6-Terra reaches 88.0% pass@1 against Fable 5's 90.4% at 19% of the cost, and locally served, open-weight Qwen3.8-27B rises from 69.2% to 92.4%, slightly exceeding Fable 5. Gains are not universal: some models are unchanged or worse. Transcript analysis attributes the gain to decomposition and persistent context. Inference-time organization can approach frontier coding accuracy at a fraction of the cost, or slightly exceed it on self-hostable weights.\n\nNeeds [uv](https://docs.astral.sh/uv/) and an API key for the model you want to test.\n\n```\ncd codebase/v2-current\nexport OPENAI_API_KEY=...\n\nLCB_RELEASE=release_v6 \\\nESCALATION_CLOUD_MAX_TOKENS=128000 \\\nESCALATION_CLOUD_TIMEOUT=7200 \\\nMULTIAGENT_MODEL=openai:gpt-5.6-terra \\\nuv run --no-project --python 3.12 --with 'datasets<4' --with numpy --with anthropic \\\n  python escalation/run_bench.py --engine multiagent --only lcb --lcb 100 --parallel 8\n```\n\n- `--engine multiagent` runs the manager;`--engine single` is the one-call baseline.\n- Other models: `anthropic:<model>` ,`dashscope:<model>` ,`openrouter:<model>` , each with its own`*_API_KEY` .\n- The pass@1 score prints at the end. Results are written to `runs/results.json` , workspaces to`runs/ws/` .\n\nCode is under the [MIT License](/slee-persis/GVS5H/blob/master/LICENSE). The paper, figures and run data are under\n[CC BY 4.0](/slee-persis/GVS5H/blob/master/LICENSE-CC-BY-4.0). The LiveCodeBench fork, the benchmark problem statements\nand the LaTeX template files keep their own licenses. See [NOTICE.md](/slee-persis/GVS5H/blob/master/NOTICE.md) for\nwhich license covers which path.", "url": "https://wpnews.pro/news/gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard", "canonical_source": "https://github.com/slee-persis/GVS5H", "published_at": "2026-09-12 13:10:54+00:00", "updated_at": "2026-09-12 13:41:33.381299+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-research", "ai-tools", "developer-tools"], "entities": ["Qwen3.8-27B", "Claude Fable 5", "GPT-5.6-Terra", "LiveCodeBench", "GVS5H"], "alternates": {"html": "https://wpnews.pro/news/gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard", "markdown": "https://wpnews.pro/news/gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard.md", "text": "https://wpnews.pro/news/gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard.txt", "jsonld": "https://wpnews.pro/news/gvs5h-five-qwen3-8-27b-models-match-claude-fable-5-on-livecodebench-hard.jsonld"}}