French Qwen3 8B
A French-optimized version of Qwen3 8B, designed primarily for autonomous development agents, DevOps workflows, and tool-using AI agents.
Features
- 🇫🇷 Responses exclusively in French
- 💻 Optimized for software development and DevOps tasks
- 🧠 16K context window
- 🎯 Low temperature (
0.2) for consistent and deterministic responses
- ⚡ Concise responses to reduce token consumption
- 🛠️ Native tool calling through Ollama
- 🤖 Designed for autonomous agents such as OpenClaw
- 📉 RTK-oriented command execution to reduce shell output and context consumption
- 🔧 Suitable for Linux, Git, Docker and infrastructure automation
Use cases
- Software development
- Bash and Linux administration
- Git and Git workflows
- Docker
- DevOps
- Infrastructure automation
- System troubleshooting
- Configuration management
- Autonomous coding agents
- Tool-based AI agents
Model
- Base model:
qwen3:8b
- Parameters: ~8B
- Quantization: Q4_K_M
- Context: 16384 tokens
- Temperature:
0.2
- Language: French
- Runtime: Ollama
- Tool calling: Native Ollama tool calls
Agent behavior
The model is configured as an autonomous development agent for OpenClaw.
It is instructed to:
- Act directly instead of unnecessarily describing its actions
- Use available tools instead of simulating their results
- Inspect only the relevant area before making changes
- Perform a targeted verification after modifications
- Retry failed operations up to two times
- Request confirmation only for genuinely destructive operations
- Never expose secrets, tokens, credentials or sensitive files
- Follow
USER.md, SOUL.md and AGENTS.md
- Reuse relevant context available to the agent
Tool calling
The model uses the native tool-calling mechanism provided by Ollama.
Tool calls are handled through Ollama’s /api/chat interface and exposed to the agent as structured tool calls rather than plain text.
This allows OpenClaw to execute tools such as:
exec- File operations
- Development tools
- Infrastructure commands
- Other tools exposed through the OpenClaw runtime
The model should never simulate a tool result. When a tool is required, it should call the appropriate tool and use the returned result.
Token efficiency
The model is configured to minimize unnecessary output and context consumption.
RTK
When RTK is available, it should be used systematically for shell commands in order to reduce command output.
The agent should:
- Prefer RTK for shell commands.
- Keep command output to the strict minimum required.
- Avoid repeating information already available in context.
- Avoid unnecessary explanations.
- Return only the required code when a coding task explicitly requires code output.
If RTK is unavailable or unusable, the agent should notify the user before falling back to standard shell commands.
Configuration
The model uses a custom Ollama Modelfile based on the official qwen3:8b model.
The official Qwen3 template is intentionally preserved to maintain compatibility with Ollama’s native tool-calling mechanism.
The custom configuration primarily defines:
- Agent behavior
- French language preference
- Token-efficient responses
- Context size
- Sampling parameters
No custom tool-calling template is required.
Requirements
- Ollama
qwen3:8b- OpenClaw or another Ollama-compatible agent runtime
- RTK (recommended for DevOps workflows)
Example
ollama create french-qwen3-8b -f Modelfile
Run the model:
ollama run french-qwen3-8b
The model can then be configured as the Ollama model used by OpenClaw.
Design goals
This model is intended to prioritize:
- Reliable tool usage
- Autonomous execution
- Concise responses
- Low context consumption
- Safe infrastructure operations
- Practical DevOps and development workflows
The objective is not to create a general-purpose conversational model, but a compact local model optimized for autonomous technical work.