The Rebellion of Local 35B

Qwen3.6 35B A3B was run with 4-bit quantization on two Tesla P40 cards to test the limits of a local coding agent.

The target was SWE-bench Verified, a benchmark that requires fixing bugs in real Python repositories to pass hidden tests.

The original text explains the starting point of the experiment: to find out why this benchmark is expensive, and whether a local model can still produce practical code patches.

As a point of comparison, it also mentions publicly disclosed figures from top-tier models, presenting context against existing performance tables such as GPT-5 Mini 56.2%.

The core point is to verify how much agentic coding performance can be pushed in a local environment with specific hardware and configuration, without relying on massive cloud models.