OxCoder-9B

Vision · Tools · Thinking · 262K context · Apache-2.0

OxCoder-9B is a lightweight coding model built for long-horizon tasks, agentic coding and agentic reasoning. It punches far above its weight class: at 9B parameters it rivals much larger models on agentic terminal and coding benchmarks.

It was distilled from frontier agent traces (Fable-5.1 and GLM-5.3 agentic coding trajectories across Claude Code, OpenCode and Codex), which gives it behaviours that matter in real iterative coding: read-before-write patterns, reacting to LSP diagnostics, and applying minimal edit diffs instead of full rewrites. It also shows unusually strong front-end reasoning for its size (UI logic, component architecture, web-native patterns).

Original model by OrionLLM · GGUF quants and vision projector by prithivMLmods · base model Qwen/Qwen3.5-9B · this Ollama packaging is a community repack, all credit to OrionLLM.

Tags

ollama run tobestyledintro/oxcoder-9b            # latest = Q4_K_M
ollama run tobestyledintro/oxcoder-9b:q5_k_m

Usage

# chat / code
ollama run tobestyledintro/oxcoder-9b

# whole-repo work (native 262144 context)
ollama run tobestyledintro/oxcoder-9b --ctx 262144

# vision (screenshots, mockups, diagrams)
ollama run tobestyledintro/oxcoder-9b "Recreate this UI as HTML/Tailwind: ./mockup.png"

API:

curl http://localhost:11434/api/chat -d '{
  "model": "tobestyledintro/oxcoder-9b",
  "messages": [{"role": "user", "content": "Refactor this function to be iterative."}]
}'

Tool calling (native, structured tool_calls — useful for agent harnesses):

tools = [{"type": "function", "function": {
    "name": "read_file",
    "description": "Read a file from disk",
    "parameters": {"type": "object", "properties": {"path": {"type": "string"}}, "required": ["path"]}}}]
r = ollama.chat(model='tobestyledintro/oxcoder-9b',
                messages=[{'role': 'user', 'content': 'Open src/main.py'}],
                tools=tools)
print(r.message.tool_calls)

Capabilities

• Coding & agentic work — trained on real agent trajectories; good at iterative edit loops, terminal tasks and front-end reconstruction.
• Tool calling — native; Ollama returns structured tool_calls (no XML parsing).
• Thinking — reasoning traces are available; keep a generous output budget (thinking can consume a few hundred tokens before the answer).
• Vision — clip vision tower (27 blocks, 1152 → 4096), shipped as the projector layer.
• Context — 262,144 tokens natively, comfortable for multi-file work.

Recommended sampling

This package ships with temperature=0.6, top_p=0.95, top_k=20.

Note the author’s own benchmark settings: agentic evaluations (Terminal-Bench 2.1, NL2Repo) use temperature=1.0, top_p=1.0; ClawEval uses temperature=0.6. If you drive it as a coding agent, raising the temperature toward 1.0 matches the way it was evaluated.

Performance

Measured on 1× RTX 5090 (32 GB), Q4_K_M fully offloaded:

Architecture

Credit & license

• Original model: OrionLLM/OxCoder-9B — Apache-2.0
• GGUF quants + vision projector: prithivMLmods/OxCoder-9B-GGUF
• Base model: Qwen/Qwen3.5-9B (Alibaba Qwen team)