flywheel-local-coder-14b

Flywheel-Local-Coder-14B

A 14.8B coding model in a single file that runs entirely on your own machine. It takes Qwen2.5-Coder-14B-Instruct and continues its pretraining on a 66-million-token corpus drawn from a real working development ecosystem, then packs the result into one Q4_K_M GGUF just under 9 GB. Your prompts and your code never leave your disk. And if you ever care to look, the whole build, corpus to weights, can be retraced hash by hash.

Run it in two commands

hf download zaindanaharper/flywheel-local-coder-14b telos-coder-14b-cpt2020-q4_k_m.gguf --local-dir .
llama-cli -m telos-coder-14b-cpt2020-q4_k_m.gguf -cnv

Prefer Ollama? Download the repo folder so the GGUF and the Modelfile sit together, then:

ollama create flywheel-local-coder-14b -f Modelfile
ollama run flywheel-local-coder-14b

No conversion step, no shards, no Python environment. The usage guide covers chat, deterministic completion, an OpenAI-compatible local endpoint, and how to verify your download against the published checksums.

Specs at a glance

Full details in the spec sheet.

What to expect

This is a local-first daily coding companion: completions, small functions, refactors, and tool-calling on your own hardware, with your code staying home.

We publish measurements, not adjectives. On our internal evaluation sets the model passes 8 of 8 baseline tasks and 8 of 10 deliberately contract-heavy hard tasks in a single attempt, with confidence intervals attached to every number. We do not claim a capability uplift over the base model: our own measurement of that difference includes zero, and the benchmarks page says so plainly. Every number there ships with the JSON it came from and the method to re-run it.

The documents

• Usage guide: run it with Ollama, llama.cpp, or as a local API.
• Benchmarks: what we measured, the intervals, and how to re-run it.
• Spec sheet: hardware guidance, training details, formats.
• Safety and claims: what this model does and does not claim.
• Model card: the full technical card.
• provenance.json and checksums.sha256: the retraceable chain from corpus to the exact bytes you downloaded.

License and attribution

Apache-2.0. Built on Qwen2.5-Coder-14B-Instruct by the Qwen team; see LICENSE for the attribution notice. The training corpus is proprietary to the author; the shipped weights carry no third-party code beyond the base model.