Bonsai 2 27B from @PrismML on my agentic coding benchmark with @CaryPalmerr’s runtime: 15/20 🔥 my old Bonsai setup scored 13/20. new setup (PTQ1_0 + his runtime) gets 2 more, with 0 loop kills Qwen3
Why CEREBRO kept it
Local model agentic coding benchmark with runtime; LLM mechanics
The text below is an automated extraction of the article at https://x.com/DogukanUrker/status/2114948826347749632, stored verbatim in the public cerebro-vault repository. Copyright remains with the original publisher (x.com).
my old Bonsai setup scored 13/20. new setup (PTQ1_0 + his runtime) gets 2 more, with 0 loop kills
Qwen3
> Local model agentic coding benchmark with runtime; LLM mechanics
Bonsai 2 27B from @PrismML on my agentic coding benchmark with @CaryPalmerr’s runtime: 15/20 🔥
my old Bonsai setup scored 13/20. new setup (PTQ1_0 + his runtime) gets 2 more, with 0 loop kills
Qwen3.8-27B official API: 16/20 Qwen3.8-27B GSQ-RCO 2-bit: 16/20 Qwen3.8-27B ByteShape 2-bit: 10/20 Qwen3.8-27B AtomicChat 2-bit: 7/20
a 5.9 GB 1.58-bit model, one task behind the official API, on a single RTX 3060
Pi Agent harness.
Backlinks
Appeared in 1 briefing