Recursive Flint on my agentic coding benchmark: 13/20 on a single RTX 3060. (config below) 20 real python bug fixes, checked by hidden tests. the model drives the Pi Agent, one repair attempt allowed
Why CEREBRO kept it
Local model benchmark on coding agent; concrete, single-source.
The text below is an automated extraction of the article at https://x.com/DogukanUrker/status/2113838778519351313, stored verbatim in the public cerebro-vault repository. Copyright remains with the original publisher (x.com).
20 real python bug fixes, checked by hidden tests. the model drives the Pi Agent, one repair attempt allowed
> Local model benchmark on coding agent; concrete, single-source.
Recursive Flint on my agentic coding benchmark: 13/20 on a single RTX 3060. (config below)
20 real python bug fixes, checked by hidden tests. the model drives the Pi Agent, one repair attempt allowed.
12 solved on the first try, +1 on repair. zero loop kills.
only 2.5B active params and a 3-bit quant, at ~49 tok/s. the 35B MoEs next to it are 4-bit and land at 15-16.
Backlinks
Appeared in 1 briefing