[Tromset]

Why

The README and Documentation.md still said the 0.3 bench was "being recorded". flycoder0.3pro also used a six-step workflow with a worked example, which had never been measured.

Measurements

All runs: 20 hidden-test problems, 4-core Linux server without a GPU, thinking off, one attempt per problem, at most 2,048 tokens per answer.

Profile	Passed	Generation
flycoder0.3 (Qwen3.5 9B)	9/20	3.9 tok/s
flycoder0.3fast (Qwen3.5 4B)	5/20	6.1 tok/s
0.2 beta (Gemma 4 12B), recorded earlier	14/20	3.4 tok/s

The 0.2 row was measured on a slower CPU, so the speeds compare only roughly.

System-prompt A/B on the same 9B base:

Prompt	Passed	Cut-off answers
FlyCoder 0.3 (shipped)	9/20	3
Same, without the final-check line	8/20	3
First pro recipe (six-step workflow + worked example)	6/20	2

What changes

• Modelfile.pro: the pro model now uses the shared prompt, plus the coding-agent rules. "Never claim you ran tests" becomes "unless a tool actually ran them". The six-step workflow and the example conversation are removed. Qwen3.8 27B, the 64K context, multi-token prediction and REQUIRES 0.32.12 stay.
• tests/modelfile.test.mjs: checks that pro keeps every shared rule and the agent rules, and has no example conversation.
• readme.md and Documentation.md: real bench tables, the prompt comparison and the reason pro was rebuilt. Limitations are updated.
• docs/bench/cpu-nothink-0.3.json: the full 0.3 run.
• docs/bench/cpu-nothink-0.3-prompts.json: the merged A/B runs.
• docs/ollama/flycoder0.3pro.md: no longer advertises the dropped workflow.

Verification

• npm test: 65 tests pass.
• Every number above comes from the committed JSON files.
• The 27B itself was not benchmarked: the server cannot hold it, so the prompt choices were measured on the 9B base.

🤖 Generated with Claude Code

https://claude.ai/code/session_013QgiejZTjqQAUv4N4EFskK

---

Generated by Claude Code