FlyCoder 0.2 beta: pure Ollama model for MacBook and Mac mini, harness removed

[Tromset]

Summary

FlyCoder becomes only the model. The 0.1 harness (CLI, web UI, Electron app, FlyBrain routing controller) is removed. FlyCoder 0.2 beta is an Ollama model with two variants, installable with one terminal command and publishable on ollama.com.

Tag	Base	Weights	Target	Context
flycoder:0.2-beta (latest)	Gemma 4 12B	7.7 GB, NVFP4 on Ollama's MLX engine	Macs with 16 GB+	32,768
flycoder:0.2-beta-fast (fast)	Qwen3.5 4B (the 0.1 base)	4.0 GB, NVFP4 on MLX	Macs with 8 GB+	16,384

Measured results

20 hidden-test coding problems. Run on a CPU-only Linux server (Ollama 0.35, GGUF Q4_K_M, thinking off as in the 0.1 harness, one sample per problem). Full answers and metrics are in docs/bench/cpu-nothink.json.

Profile	Passed	Decode	Tokens/answer	Time/problem
flycoder0.1beta (0.1)	5/20	6.6 tok/s	581	94 s
flycoder:0.2-beta-fast	5/20	6.5 tok/s	537	89 s
flycoder:0.2-beta	14/20	3.4 tok/s	739	235 s

• The default variant solves almost 3× more problems, but takes about 2.5× the time per problem on CPU. Multi-token prediction helps: 77% of 6,594 drafted tokens were accepted (3.35 tokens per verification step), and generation was +49% faster than with drafting disabled. Ollama reports about +90% on Apple Silicon, so the gap should narrow on a Mac. This is not measured yet.
• The fast variant keeps 0.1's speed and pass rate, and writes 8% fewer tokens. On a Mac it also gets the MLX engine.
• Seven default-variant problems and one fast-variant problem were rerun. On the first pass the server ran out of memory or the bench's 5-minute fetch timeout cut the request. These were infrastructure failures, not answers. The bench now streams answers and retries a crashed runner once.

Changes

• Modelfile, Modelfile.fast: the two variants. They use each base model's official sampling, set an explicit context (Ollama otherwise defaults to 4K below 24 GB), and add a short code-quality system prompt.
• install.sh: curl -fsSL https://raw.githubusercontent.com/Tromset/flycoder/main/install.sh | sh. It checks the Ollama version (0.31 or later), picks the variant from the Mac's memory, and falls back to GGUF bases on Intel Macs, Linux and Windows.
• scripts/publish.sh <user>: pushes 0.2-beta, latest, 0.2-beta-fast and fast to ollama.com. It refuses to push non-MLX builds as the main tags.
• bench/: the benchmark runner and problems (npm run bench). On macOS it uses sandbox-exec (no network, no writes outside the scratch dir). bench/baselines/ keeps the 0.1 profile so the two versions can be compared.
• tests/ + .github/workflows/test.yml: every bench test is checked against a reference solution and must reject a stub. Modelfiles, installer and publisher are checked for consistency. CI runs on Linux and macOS.
• Harness code, its docs and its dependencies are removed (package.json now has no dependencies). The independent Qwen 2.5 fly-language QLoRA pipeline is kept.

Validation

• npm test: 52/52 locally. CI is green on ubuntu and macos; on macOS the reference solutions run inside the sandbox.
• Ollama 0.35.0: ollama create succeeds for both MLX Modelfiles. The models report tools, thinking and vision, and keep the gemma4 / qwen3.5 renderers and parsers.
• curl … | sh installer tested from this branch (GGUF path on Linux).

Not verified here

• Speed on Apple Silicon: MLX models cannot run in this Linux container. Run npm run bench on a Mac. If the default variant is too slow there, ollama cp flycoder:0.2-beta-fast flycoder:latest makes the fast one the default.
• Thinking-mode quality and length were not benchmarked.
• Publishing needs an ollama.com account and ollama signin on the Mac.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MDAB5YFn5ZSH37mcH3Tfid