GVS5H harness lets five Qwen3.8-27B agents match Claude Fable 5

GVS5H harness lets five Qwen3.8-27B agents match Claude Fable 5

A GitHub repository called GVS5H reached 58 points and 15 comments on Hacker News on 25 October 2026. Its author, GitHub user slee-persis, describes a method in which fresh instances of a single model split up a coding problem and coordinate through files written to a shared disk. The repository reports the resulting setup pushing Alibaba's open-weight Qwen3.8-27B and Qwen3.8 Flash Next to accuracy that matches or beats Anthropic's Claude Fable 5 on 100 hard LiveCodeBench problems.

GVS5H is an open-source technique called ledger-based zero-shot self-orchestration. A manager and several fresh-context copies of the same model split a coding problem and share a plan, notes and a running solution file on disk instead of one long context.

What changed

The harness changes how a single model is used for coding. Fresh instances of the same underlying model fill every role but one, a separate verifier, each called in a new context and reading and writing to a shared workspace. task.md holds the problem statement, plan.md the manager's overall plan, tasks.json a debug mirror of the task list, notes.md curated findings, and solution.py the current answer.

On 100 of the latest hard LiveCodeBench problems, the authors report two routes to frontier-level accuracy. Orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1 against Claude Fable 5's single-call score of 90.4%, at about 9% of the cost. Locally served, open-weight Qwen3.8-27B rises from 66.8% to 92.4% pass@1 once orchestrated. Across eleven open and closed-weight models tested, the method produces gains of up to 25.6 percentage points on pinned backends, though the authors state gains are not universal and some models score the same or worse. A comparison figure also includes Opus 5, with a caption noting Opus 5 ran on an earlier version of the scaffold, in a single pass served through OpenRouter rather than with five calls on a pinned backend, so that result is not directly comparable to the rest.

The code is MIT licensed; the paper, figures and run data are CC BY 4.0. By the time it reached the front page on 25 October 2026, the repository had drawn 785 stars and 94 forks. This is not GVS5H's first appearance on Hacker News: the same repository was submitted three times in September 2026 under earlier titles and drew six points or fewer each time, before this week's resubmission reached 58 points and 15 comments within its first hour.

Qwen3.8-27B itself is Alibaba's 27 billion parameter, Apache 2.0 licensed dense model, released in mid August 2026 as an update to Qwen3.6-27B with stated gains in coding. The Next Web reported on 19 August 2026 that the model passed a million downloads within days of release and that the research firm Artificial Analysis scored it level with OpenAI's GPT-5.6 Luna, while noting it is slow because it reasons at length before answering.

Why it matters for agent builders

It matters because the pattern is not tied to Qwen or to coding. A manager model that writes a plan to disk and hands pieces of it to fresh worker instances of itself is a way to try multiplying calls to a model you already have before paying for a larger one. The shared-file coordination, a plan file, a notes file, a running solution file, is simple enough to copy into an existing agent harness without adopting the whole project.

Matching a frontier model's score at roughly a tenth of the per-task cost changes the math for any team running coding agents at volume, if the result holds up outside this one benchmark.

One Hacker News commenter who said they already run Qwen3.8-27B for coding described doing something close to this by hand: a plan file, a status file and a notes file, with no formal harness driving it, just manual turn-taking. That suggests the file-based planning pattern was already emerging informally among people running local models before this project named it.

What to watch

Several commenters on the Hacker News thread pushed back on what the benchmark shows. One, posting as vikramkr, called the comparison "meaningless," pointing out there is no comparison against newer frontier models (Astra, GPT 6.1 Sol, Fable 5.1 or Opus 5.5) and no test on agentic, tool-using work, which is what current models are optimized for; they also flagged that LiveCodeBench Hard's problems are old enough that memorization, not reasoning, could explain some of the score. Another commenter, posting as sams99, objected that the baseline Claude Fable 5 score comes from a single-shot mode with no tools, while the orchestrated Qwen score comes from a multi-call setup with delegation, which they argued is not a fair comparison. Whether the gains GVS5H reports survive a tool-using, contamination-resistant benchmark is a question the project has not yet answered.

Key takeaways

• GVS5H is an open-source, MIT-licensed method (ledger-based zero-shot self-orchestration) that has fresh instances of one model coordinate through plan, notes and solution files on a shared disk.
• On 100 hard LiveCodeBench problems, orchestrated Qwen3.8 Flash Next reportedly reaches 93.0% pass@1 versus Claude Fable 5's 90.4%, at about 9% of the cost; locally served Qwen3.8-27B rises from 66.8% to 92.4%.
• The authors say gains are not universal: some of the eleven models tested showed no improvement or a worse score under orchestration.
• Hacker News commenters challenged the benchmark choice itself, citing possible contamination in LiveCodeBench Hard and the lack of any agentic, tool-calling comparison.
• Qwen3.8-27B is Alibaba's 27 billion parameter open-weight model, released in mid August 2026, which The Next Web reported passed a million downloads within days.

FAQ

What is GVS5H?

GVS5H is an open-source GitHub project that uses ledger-based zero-shot self-orchestration: a manager and several fresh-context copies of one model split a coding problem and coordinate by reading and writing a plan file, a notes file and a solution file on a shared disk, rather than relying on one long conversation.

How much cheaper is the orchestrated Qwen3.8 Flash Next than Claude Fable 5?

The project's authors report that orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1 on 100 hard LiveCodeBench problems, slightly above Claude Fable 5's single-call score of 90.4%, while costing about 9% as much per task on a pinned backend.

Does the benchmark reflect real agentic, tool-using work?

Not directly. The comparison runs on LiveCodeBench Hard, a code-generation benchmark with no tool calls, and several Hacker News commenters argued that today's models are optimized for agentic tool use rather than single-shot code generation, and that the benchmark's age raises a risk of memorization rather than genuine reasoning gains.

Source: GVS5H on GitHub Discussion: Hacker News

Compare coding-focused agents and harnesses in the coding category on AI Agents Listing.