Sushi Coder on Linux: A Complete Setup Guide

Most local AI coding tools feel like a compromise. This one doesn’t — here’s how to run a reinforcement-learning-trained coding model entirely on your own hardware, including older AMD GPUs that ROCm officially ignores.

What Is Sushi Coder?

Qwen3.5-9b-Sushi-Coder-RL is a fine-tuned, reinforcement-learning-trained coding model built on top of Qwen 3.5 9B. It was trained on:

• Claude Opus 4.6 reasoning traces (nohurry/Opus-4.6-Reasoning-3000x-filtered)
• Competitive programming problems and solutions (open-r1/codeforces-cots)

After supervised fine-tuning, a reinforcement learning stage was run specifically for coding using NousResearch’s Atropos pipeline. The result is a model that reasons through code problems rather than pattern-matching its way through them.

This guide covers the GGUF version, which runs on any Linux machine via Ollama — with optional AMD GPU acceleration for older unsupported cards.

---

Prerequisites

• Linux (Ubuntu 22.04+ recommended)
• Python 3.10+
• At least 8GB RAM (16GB+ recommended)
• Optional: AMD GPU with RDNA2 architecture (RX 6000 series)

---

Step 1 — Install Ollama

Ollama manages local model serving and exposes an OpenAI-compatible API endpoint.

curl -fsSL https://ollama.com/install.sh | sh

Verify it’s running:

ollama --version
systemctl status ollama

---

Step 2 — Download the Model

~5.6 GB download — make sure you have enough space and a stable connection.

Install the Hugging Face CLI if you don’t have it:

pip3 install huggingface_hub --user
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

Download the Q4_K_M quant (5.63 GB — best balance of speed and quality):

hf download bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF \
  Qwen3.5-9b-Sushi-Coder-RL.Q4_K_M.gguf

---

Step 3 — Register the Model with Ollama

Create a Modelfile pointing at your downloaded GGUF:

GGUF_PATH=$(find ~/.cache/huggingface/hub -name "Qwen3.5-9b-Sushi-Coder-RL.Q4_K_M.gguf" | head -1)
echo "FROM $GGUF_PATH" > ~/Modelfile

cat >> ~/Modelfile << 'EOF'

PARAMETER num_ctx 16384

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

SYSTEM "You are a helpful coding assistant."
EOF

Register it with Ollama:

ollama create sushi-coder -f ~/Modelfile

Test it on CPU first to confirm it works:

ollama run sushi-coder "Write a Python hello world"

You should see a response generated. Once confirmed, move on.

---

Step 4 — Enable AMD GPU Acceleration (Optional - for RX 6750XT 12GB users)

⚠️ Save your work before attempting this. This voids your warranty and could damage your system. I do not recommend it.

Skip this section if you don’t have an AMD GPU, or if your GPU is officially ROCm-supported. Only follow this section for older RDNA2 cards like the RX 6700 XT that aren’t on AMD’s official support list. GPU driver issues can crash systems — test CPU mode works first and keep the revert command ready.

The trick is to override the GPU version so ROCm treats your card as a supported RDNA2 GPU, without installing the full ROCm stack system-wide (which is the common cause of crashes).

Create a systemd override file:

sudo mkdir -p /etc/systemd/system/ollama.service.d
echo '[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"' | sudo tee /etc/systemd/system/ollama.service.d/override.conf

Verify it saved correctly:

cat /etc/systemd/system/ollama.service.d/override.conf

Should show:

[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=10.3.0"

Reload and restart Ollama:

sudo systemctl daemon-reload
sudo systemctl restart ollama

Test again — if it’s noticeably faster, the GPU is working:

ollama run sushi-coder "Write a Python quicksort"

To monitor GPU activity while generating:

sudo apt install radeontop
radeontop

To revert if anything goes wrong:

sudo systemctl revert ollama
sudo systemctl restart ollama

---

Step 5 — Install OpenCode

OpenCode is an open-source agentic coding tool — the closest open-source equivalent to Claude Code. It can read files, write files, run shell commands, and work through multi-step coding tasks.

curl -fsSL https://opencode.ai/install | bash

---

Step 6 — Configure OpenCode for Sushi Coder

OpenCode needs to know about your local Ollama server. Create the config:

mkdir -p ~/.config/opencode
cat > ~/.config/opencode/opencode.json << 'EOF'
{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "local-ollama": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Local Ollama",
      "options": {
        "baseURL": "http://localhost:11434/v1"
      },
      "models": {
        "sushi-coder": {
          "name": "Sushi Coder",
          "options": {
            "num_ctx": 32768
          }
        }
      }
    }
  }
}
EOF

Why local-ollama and not ollama? Using a custom provider ID prevents OpenCode from confusing your model with the built-in Ollama registry, which would trigger a tool support check and reject the model.

---

Step 7 — Set Up a Project and Start Coding

Create an isolated directory for your project:

mkdir -p ~/projects/my-project
cd ~/projects/my-project

Launch OpenCode:

opencode

Inside OpenCode, select your model by typing /models and choosing Sushi Coder under the Local Ollama provider.

You’re now running a locally fine-tuned coding AI with full agentic capabilities — file reading, file writing, shell access — entirely on your own hardware.

---

Quant Comparison

Quant	Size	VRAM	CPU RAM needed	Quality	Speed
Q4_K_M	5.63 GB	~7 GB	~8 GB	Good	Fast
Q8_0	9.53 GB	~11 GB	~12 GB	Better	Slower

For most coding tasks Q4_K_M is the right choice. Use Q8_0 if you have 12GB+ VRAM and want higher quality responses.

---

Troubleshooting

huggingface-cli: command not found Use hf instead — it’s the same tool with a shorter alias.

Ollama not finding the model Run ollama list to confirm sushi-coder appears. If not, re-run ollama create sushi-coder -f ~/Modelfile.

OpenCode says model doesn’t support tools Make sure your provider ID in opencode.json is local-ollama (not ollama). The built-in ollama provider does a registry check that rejects custom models.

GPU crashes system Run sudo systemctl revert ollama && sudo systemctl restart ollama to fall back to CPU. CPU mode works fine for interactive coding use.

Slow generation First response is always slow (model load). Subsequent responses are faster. On CPU expect ~10-20 tokens/sec on a modern desktop CPU.

---

Resources

• Model: https://huggingface.co/bigatuna/Qwen3.5-9b-Sushi-Coder-RL-GGUF
• Ollama: https://ollama.com
• OpenCode: https://opencode.ai
• Training pipeline: https://github.com/NousResearch/atropos