FlyCoder 0.3

FlyCoder is a local coding model for Ollama, tuned for MacBooks and Mac minis: complete, runnable code, a large context, answers in your language.

ollama run Tromset/flycoder0.3       # Macs with 16 GB
ollama run Tromset/flycoder0.3fast   # Macs with 8 GB
ollama run Tromset/flycoder0.3pro    # Macs with 32 GB or more

You need Ollama 0.30 or later (0.32.12 for pro). Full documentation: Documentation.md.

The three models

Model	Base	Weights	Mac	Context
Tromset/flycoder0.3	Qwen3.5 9B	6.6 GB	16 GB	65,536 tokens
Tromset/flycoder0.3fast	Qwen3.5 4B	3.3 GB	8 GB	32,768 tokens
Tromset/flycoder0.3pro	Qwen3.8 27B	18 GB	32 GB or more	65,536 tokens

All three are portable GGUF builds: they run on Apple Silicon and Intel Macs, Linux and Windows, with tool calling, thinking and images. The bases are licensed under Apache 2.0.

What FlyCoder sets up on top of the base:

• A large context: 64K tokens instead of Ollama's 4,096 default on machines under 24 GB. Qwen's hybrid attention keeps it cheap: flycoder0.3 uses 8.8 GB in total at 64K.
• The right sampling: Qwen's recommended settings for coding in thinking mode.
• A short system prompt for code: exact names and signatures, complete code without "TODO", edge cases, no invented APIs, a final check before answering, secure defaults, answers in the user's language.
• flycoder0.3pro adds rules for coding agents and multi-token prediction on Qwen3.8 27B. Details: Documentation.md, section 6.5.

Install under a short name

curl -fsSL https://raw.githubusercontent.com/Tromset/flycoder/main/install.sh | sh
ollama run flycoder

The installer picks the variant from your memory and names it flycoder. Options after sh -s --: --fast, --pro, --all, --mlx (Ollama's MLX engine on Apple Silicon), --no-brain.

Measured quality and speed

20 coding problems with hidden tests (bench/), on a 4-core Linux server without a GPU, thinking off, one attempt per problem:

Profile	Base	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 (did not run on every 16 GB Mac)	Gemma 4 12B	14/20	3.4 tok/s

Three of the 9B's failures are answers cut off at 2,048 tokens: with thinking off it reasons in code comments; keep thinking on (the default) for code you ship. On a Mac, speeds are much higher. Details and the system-prompt comparison: Documentation.md, section 7.

FlyBrain: less RAM for simple requests (optional)

FlyBrain is a small Node server in front of Ollama. It answers to flycoder0.3 and flycoder0.3fast, sends simple questions to a smaller expert and hard requests (and every agent request with tools) to the strong one, and keeps a single expert in memory. On 60 labelled requests it routes 59 correctly, and no hard request goes to the small model.

node ~/.flycoder/brain/flybrain.mjs        # installed by install.sh; or: npm run brain
OLLAMA_HOST=127.0.0.1:11435 ollama run flycoder0.3

Details: Documentation.md, section 5.5.

Publishing

Every change to main that touches a Modelfile or a page in docs/ollama/ is published on ollama.com by the publish-ollama workflow: models first, then their pages. It needs two repository secrets, OLLAMA_KEY and OLLAMA_API_KEY (Documentation.md, section 8).

Each publication is then mirrored on Hugging Face by the publish-huggingface workflow: the same GGUF weights, vision projector, system prompt and sampling settings, with a model card from docs/huggingface/, runnable with ollama run hf.co/<namespace>/flycoder0.3 or llama.cpp. It needs the HF_TOKEN secret (Documentation.md, section 8.1).

Repository contents

• Modelfile, Modelfile.fast, Modelfile.pro: the three variants. Modelfile.lite, Modelfile.router: FlyBrain's experts.
• docs/ollama/: the ollama.com page of each model. docs/huggingface/: the Hugging Face model card of each model.
• brain/: the FlyBrain server and its routing rules.
• install.sh: one-command install. scripts/publish.sh, scripts/update-ollama-pages.sh, scripts/publish-hf.sh: publishing.
• bench/: quality and speed benchmark. tests/: npm test.
• training/, scripts/build_language_dataset.js, server/fly-language.js: the former FlyBrain voice pipeline, independent of FlyCoder.

License

Repository code: MIT, see license.md. The weights keep the license of their base: Apache 2.0 for Qwen3.5 and Qwen3.8.