Five Qwen3.8-27B models match Fable 5 on hard coding problems, if you read the benchmark kindly
A repo called GVS5H lifts a local 27B model from 66.8% to 92.4% pass@1 on LiveCodeBench Hard by orchestrating copies of it through a shared folder.
The method needs no training: a manager instance writes a plan to disk and fresh worker instances pick up pieces.
A commenter's objection: the benchmark is old enough that contamination is likely.
What the harness does
GVS5H ("Gorilla vs. 5 Humans") is a training-free scheme in which fresh instances of one model split a problem and coordinate through a shared filesystem holding a plan, notes and the current solution. On the 100 latest hard LiveCodeBench problems, the authors report Qwen3.8-27B rising from 66.8% to 92.4% pass@1 when run locally. Another Qwen model reaches 93.0% against 90.4% for Fable 5, at about 9% of the cost. Some models gain nothing or get worse, and the Opus 5 comparison ran on an earlier scaffold.
The thread's verdict on the evidence
"No comparison to astra/6.1 sol/fable 5.1/opus 5.5. No comparison on a benchmark that reflects actual agentic work… essentially guarenteed contamination." — vikramkr, who read the benchmark's dataset description
Another reader went to the "10x the time" question: the cost saving is real only if you accept a slower result. A third commenter pointed at harness quality, saying that getting Opus 5.5 and Fable 5.1 to spin up non-blocking sub-agents inside Claude Code took four or five attempts.
If you run local models, the transferable idea is the shared-folder ledger. The 92.4% is a number to rerun on your own tasks first.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
France asked its telecom operators to be ready to block TikTok, Instagram, Snapchat and X
Blast says it read a letter from the prime minister's office to Orange, SFR, Free and Bouygues ahead of the 24 October demonstration.