AdamBench ranks local coding LLMs

This feels less like a universal leaderboard and more like a brutally honest local-model reality check, which is exactly why it’s useful. The score favors not just raw coding quality, but also how well a model survives iterative repair loops without wasting time or tokens.

• –Qwen3.5 122b A10b takes the top AdamBench score, but Qwen3.5 35b A3b is the author’s daily driver because it balances quality, speed, and context headroom.
• –gpt-oss-120b and gpt-oss-20b are the standout surprises: fast for their size and unusually token-efficient, which matters a lot in agentic coding.
• –Nemotron models lag hard on quality and efficiency; even the best one lands around the top 10, with huge reasoning-token overhead.
• –The benchmark is intentionally single-run and self-repair heavy, so it measures real workflow resilience more than clean one-shot coding ability.
• –Models that failed on tool calling or chat templates were excluded, which is harsh but sensible if the goal is practical local usability.