A lot of the recent v0.5.0 work was about hardening coding-task behavior, improving validation and completion behavior, and reducing the ways evals can be gamed.

One thing that stood out during that work was OmniCoder-9B Q8_0. I care less about “looks good in a demo” and more about whether a small model still holds up under real repo tasks, explicit validation, replayable runs, and stricter task contracts without relying on benchmark-specific hacks.

OmniCoder-9B Q8_0 is one of the few small local models that has felt genuinely solid in that environment for me. I am not making a broad “best model” claim, but it stayed on track better under constraints that usually expose weak reasoning, fake progress, or contamination.

Curious if others doing real local coding workflows have seen similar results with OmniCoder or other small models.