OmniCoder-9B

A 9B coding agent fine-tuned on 425K agentic trajectories.

Get Started | Benchmarks | GGUF Downloads

Overview

OmniCoder-9B is a 9-billion parameter coding agent model built by Tesslate, fine-tuned on top of Qwen3.5-9B's hybrid architecture (Gated Delta Networks interleaved with standard attention). It was trained on 425,000+ curated agentic coding trajectories spanning real-world software engineering tasks, tool use, terminal operations, and multi-step reasoning.

The training data was specifically built from Claude Opus 4.6 agentic and coding reasoning traces, targeting scaffolding patterns from Claude Code, OpenCode, Codex, and Droid. The dataset includes successful trajectories from models like Claude Opus 4.6, GPT-5.4, GPT-5.3-Codex, and Gemini 3.1 Pro.

The model shows strong agentic behavior: it recovers from errors (read-before-write), responds to LSP diagnostics, and uses proper edit diffs instead of full rewrites. These patterns were learned directly from the real-world agent trajectories it was trained on.

Key Features

Trained on Frontier Agent Traces : Built from Claude Opus 4.6, GPT-5.3-Codex, GPT-5.4, and Gemini 3.1 Pro agentic coding trajectories across Claude Code, OpenCode, Codex, and Droid scaffolding
Hybrid Architecture : Inherits Qwen3.5's Gated Delta Networks interleaved with standard attention for efficient long-context processing
262K Native Context : Full 262,144 token context window, extensible to 1M+
Error Recovery : Learns read-before-write patterns, responds to LSP diagnostics, and applies minimal edit diffs instead of full rewrites
Thinking Mode : Supports <think>...</think> reasoning chains for complex problem decomposition
Apache 2.0 : Fully open weights, no restrictions

Benchmarks

Benchmark	OmniCoder-9B	Qwen3.5-9B	Qwen3-Next-80B	GPT-OSS-120B	GPT-OSS-20B	GLM-4.7-Flash	GLM 4.7	Claude Haiku 4.5
AIME 2025 (pass@5)	90				91.7	91.6
GPQA Diamond (pass@1)	83.8	81.7	77.2	80.1	71.5			73
GPQA Diamond (pass@3)	86.4
Terminal-Bench 2.0	28.1	20					33.4	27

GPQA Diamond pass@1: 83.8 (166/198). +2.1 points over the Qwen3.5-9B base model (81.7). At pass@3: 86.4 (171/198).
AIME 2025 pass@5: 90 (27/30).
Terminal-Bench 2.0: 28.1 (25/89). +8.1 points over the Qwen3.5-9B base model (20).

Quickstart

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Tesslate/OmniCoder-9B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": "You are a helpful coding assistant."},
    {"role": "user", "content": "Write a Python function to find the longest common subsequence of two strings."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.95, top_k=20)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

vLLM

vllm serve Tesslate/OmniCoder-9B --tensor-parallel-size 1 --max-model-len 65536

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="token")
response = client.chat.completions.create(
    model="Tesslate/OmniCoder-9B",
    messages=[{"role": "user", "content": "Explain the difference between a mutex and a semaphore."}],
    temperature=0.6,
)
print(response.choices[0].message.content)

llama.cpp (GGUF)

llama-cli --hf-repo Tesslate/OmniCoder-9B-GGUF --hf-file omnicoder-9b-q4_k_m.gguf -p "Your prompt" -c 8192

All quantizations: Tesslate/OmniCoder-9B-GGUF

Training Details


Base Model	Qwen3.5-9B
Method	LoRA SFT (r=64, alpha=32)
Dataset	425K agentic trajectories from 5 sources
Packing	Sample packing with 99.35% efficiency
Hardware	4x NVIDIA H200 (DDP)
Framework	Axolotl
Precision	bf16
Optimizer	AdamW (lr=2e-4, cosine schedule)

Architecture

OmniCoder inherits Qwen3.5-9B's hybrid architecture:

Gated Delta Networks : Linear attention layers interleaved with standard attention for efficient long-range dependencies
VLM Backbone : Built on Qwen3_5ForConditionalGeneration

Recommended Sampling Parameters

Parameter	Value
Temperature	0.6
Top-P	0.95
Top-K	20
Presence Penalty	0.0

For agentic / tool-calling tasks, consider lower temperature (0.2-0.4) for more deterministic behavior.

Limitations

Performance on non-English tasks has not been extensively evaluated
Tool-calling format is flexible but works best with the scaffolding patterns seen in training

Acknowledgments

Special thanks to the Axolotl team and the discussion in axolotl#3453 for helping get Qwen3.5 packing support working.

Citation

@misc{omnicoder2025,
  title={OmniCoder-9B: A Frontier Open Coding Agent},
  author={Tesslate},
  year={2025},
  url={https://huggingface.co/Tesslate/OmniCoder-9B}
}

Built by Tesslate