carstenuhlig/ omnicoder-2-9b:q4_k_m

1,618 1 week ago

2nd gen OmniCoder, fine-tuned from Qwen3.5-9B. Trains on assistant tokens only (unlike v1): no more repetition loops, stable tool-calling in long agentic sessions. Original: https://hf.co/Tesslate/OmniCoder-2-9B

tools thinking 
ollama run carstenuhlig/omnicoder-2-9b:q4_k_m

Details

1 week ago

70edfef80929 · 5.7GB ·

model
qwen35
·
8.95B
·
Q4_K_M
system
You are OmniCoder 2, an expert coding assistant trained on agentic software engineering trajectories
params
{ "num_ctx": 32768, "repeat_penalty": 1, "stop": [ "<|im_start|>", "<|im
template
{{- $lastUserIdx := -1 -}} {{- range $idx, $msg := .Messages -}} {{- if eq $msg.Role "user" }}{{ $la

Readme

OmniCoder 2 9B

Fine-tune of Qwen3.5-9B on 425K agentic coding trajectories: terminal agent runs, SWE-bench patches, tool-use sequences. Built for IDE coding agents (OpenCode, Cline, Roo Code) and terminal pipelines, not general chat.

v2 trains on assistant tokens only. v1 saw all tokens including template boilerplate, which caused repetition loops and unstable tool-calling in long sessions. v2 also preserves think blocks on every turn, so the model reasons throughout a multi-step session rather than just at the final answer.

Original model: Tesslate/OmniCoder-2-9B

Benchmarks (self-reported)

Benchmark OmniCoder 2 9B Base Qwen3.5-9B
Terminal-Bench 2.0 25.8% 14.6%
GPQA Diamond pass@1 83% 81.7%
AIME 2025 pass@5 90 91.6

Parameters and quantization

Temperature 0.6 (0.2-0.4 for tool-heavy agentic use), top-p 0.95, top-k 20, context 32768.

Quant Size Notes
Q4_K_M (this) 5.7 GB 8 GB VRAM, recommended
Q5_K_M 6.5 GB Better tool-call reliability
Q8_0 9.5 GB Near-lossless
BF16 17.9 GB Best for production pipelines

I know i have not yet added those 3 other quantizations

At Q4_K_M and below, tool-call failures increase in long agentic loops. Lower temperature or step up to Q5_K_M helps.