Blog
Benchmarks, findings, and notes from building Polyglot.
October 1, 2026
Six coding agents, seven local models, 30 runs each
A larger, stricter rerun of our tool-call reliability benchmark - seven self-hosted models, six agents, tiers fixed in advance, and every run public.
September 28, 2026
Is it just a chat template? We checked.
A follow-up on open-weight tool-call reliability - ruling out chat template misconfiguration with a direct diff against the model's own official template, raw failing examples, a cross-backend vLLM test, and what actually does fix it.
September 28, 2026
We tested the most current model - and found two bugs in our own code along the way
Testing tool-call reliability on a current-generation model (Qwen3.8-27B) - a genuinely different result than anything tested before, and two real bugs it surfaced in Polyglot itself.
September 22, 2026
Your local coding agent might be doing nothing at all
A head-to-head on open-weight tool-call reliability, and why Polyglot parses tool calls out of text instead of trusting the model to format them.