Five Small Models Team Up and Match the Strongest Coding Model
GVS5H's method is called ledger-style zero-shot self-orchestration: no extra training, with multiple fresh model instances collaborating through a shared file system holding plans, notes and the current solution. The test used the latest 100 hard problems from LiveCodeBench, each model run five times with reasoning enabled.
- 01THE PROJECT — On October 25, GVS5H reached the Hacker News front page and its ledger-style self-orchestration method went public.
- 02THE SETUP — Five Qwen instances collaborate through a shared file system, with plans, notes and solutions all on disk.
- 03THE RESULT — Qwen3.8-27B rises from 66.8% to 92.4% and Flash-Next reaches 93.0%, both above Fable 5's single-call score.
WHAT HAPPENEDWhat Happened
The numbers come from the project team and the Hacker News discussion: a locally deployed open-source Qwen3.8-27B scored 66.8% on its own, rising to 92.4% when orchestrated, a gain of 25.6 percentage points that edges past Claude Fable 5's 90.4% in a single call. The larger Qwen3.8-Flash-Next reached 93.0% orchestrated, at what the project team says is 9% of Fable 5's cost.
The method is not universal: among 11 models tested, some saw no improvement or got worse; the loop itself consumes more compute than a single call, and the project offers no equal-token comparison. The benchmark is also contested: problems designed for single-turn evaluation wrapped in an agent loop are not quite the same contest, and LiveCodeBench is a 2025 benchmark that some researchers consider saturated.
THE METHODThe Ledger Matters More Than the Model
The core of GVS5H is not a new model but a collaboration ledger: plan files, task lists and verification roles, all written into a shared file system. Model instances never talk to each other directly; they divide the work by reading and writing the ledger.
The project team's attribution is direct: the gains come from task decomposition and persistent context, not from any instance suddenly getting smarter. How inference is organised is becoming a performance variable on par with training.
THE COST QUESTIONCheap, Relative to a Flagship
Flash-Next at 93.0% orchestrated for about 9% of Fable 5's cost reads easily as a total victory. The real ledger is more complicated: looped calls consume more tokens, and local deployment carries its own hardware and operations costs.
The true cost comparison is not the price of a single call but the total spend to finish one hard problem. The project offers no equal-token comparison, so the 9% figure should be read as the project's own accounting, not a procurement conclusion.
THE BENCHMARK CAVEATSame Contest, Different Rules
The dispute belongs in the main text too: LiveCodeBench was designed for single-turn evaluation, and wrapping it in an agent loop changes the rules; the benchmark itself is a 2025 problem set that some researchers consider saturated.
That does not negate the method's value, only the scope of the conclusion: orchestration demonstrates the potential of inference-time organisation, not that Qwen3.8-27B beats flagship models across all coding tasks.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISFormation Sometimes Beats Physique
The value of GVS5H is not the 92.4% figure but what it proves: inference-time organisation can close a training-scale gap. Tokenta Signal's judgment is unchanged: the next leg of the coding race is an orchestration race, and what model vendors will sell is not just weights, but the method for arranging small models into a flagship formation.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The original report confirms the core figures: Qwen3.8-27B rising from 66.8% to 92.4%, Fable 5 at 90.4%, Flash-Next at 93.0% orchestrated at roughly 9% of the cost, while noting both the non-universality caveat and the benchmark dispute.
PRIMARY SOURCESPrimary Sources
- PakTechStan: Five Small Open Models Just Outscored Claude's Best Coder
https://paktechstan.com/five-small-open-models-just-outscored-claude-s-best-coder-with-no-extra-training - GitHub project GVS5H
https://github.com/slee-persis/GVS5H - ai0.news daily digest, 2026-10-25: Cerebras Spans 59 Wafers, Five Qwens Tie Claude Fable 5
https://ai0.news/posts/2026-10-25-daily-digest/
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook