Project · on Hugging Face
Sharp Models: local coders that get to the point
Open-weight models on Hugging Face, tuned to make local agentic coders stronger, faster and leaner with tokens: TielCoder-35B-A3B, Dirk-27B and SharpSpark-4B.
peculiar-ragdoll
What
I publish local models on Hugging Face: over six million downloads so far, across thirty repos. They aren’t new models trained from scratch. Each takes a strong open model and makes it a better agentic coder on your own machine: it solves as much or more, in less time, with fewer tokens.
The three most downloaded families:
- TielCoder-35B-A3B, the fast one. A mixture-of-experts coder built on Ornith-1.5: at 4-bit and 23 GB it fixes 12 of 25 real repository issues on SWE-bench-Live, level with Opus 4.6 (medium), and runs well even when it has to spill from the GPU into system RAM. CyberTiel is its uncensored sibling, for security work.
- Dirk-27B, the thorough one. Qwen3.8-27B with the Sharp template: the same weights, asked better. It solves 15 of the same 25, one fewer than stock, and reaches a fix 2.7 times faster.
- SharpSpark-4B, the small one. A long-context 4B coder that fits a 6 GB GPU, tied with Haiku 4.5 (high) on a slice of SWE-bench-Live, and spending 43% fewer tokens per question than its stock version at the median, at the same accuracy.
How
Three levers, used where each one helps:
- Chat template fixes. A model’s chat template decides how it sees a conversation, and a bad one quietly costs you. The Sharp chat templates fix Qwen’s (keeping the model’s earlier reasoning, tool calls that parse) and add the Sharp system prompt: lead with the answer, cut the filler, keep what correctness needs.
- Importance matrices. Quantizing works best when it knows which weights the work leans on. Mine are calibrated on curated text weighted toward agentic coding and security, at lengths up to 262k tokens.
- Dynamic quantization. More bits where they matter and fewer where they don’t, built on the best existing quant recipes and keeping the multi-token-prediction head for faster generation. SharpQuant is where I’m taking this next.
Measured like work
The builds are measured on what they’re for: real repository fixes on SWE-bench-Live (judged by the projects’ own tests, not another model), multi-turn conversations on Claw-Eval, and knowledge on MMLU-Pro. And the model cards say what each build costs as well as what it gains: TielCoder is a coder, and cheerfully bad at trivia.