[bmdpat]
Download BMD free

Agent checks · measured 2026-10-12

Can your local model do agent work?

Can a local model do small coding tasks for an agent? Each model below did the same checked tasks through Codex CLI on an RTX 5090.

Codex CLI in OSS mode drove each model through Ollama in a container. Each task's check was written before the run and ran after the agent stopped, with run_check.py and tests/ restored from the fixture first.

  1. gemma4:26b · Q4_K_M10 of 10 checked coding tasks passed
  2. qwen3.5:9b · Q4_K_M5 of 10 checked coding tasks passed
  3. llama3.1:8b · Q4_K_M0 of 10 checked coding tasks passed
  4. gpt-oss:20b · MXFP40 of 10 checked coding tasks passedOllama closed the stream in 10 of 10

Find the model that fits your GPU first: open the sizing desk.