gemma-4-opencoder-48b-a10b-nvfp4

Gemma 4 OpenCoder 48B-A10B

Gemma 4 OpenCoder 48B-A10B is a post-trained Gemma 4 coding model focused on OpenCode-style software engineering workflows, long-context reasoning, and tool use.

The model has approximately 48.1B total text parameters and approximately 9.5B active text parameters per token. It supports a 262K token context window through the underlying Gemma 4 configuration.

This NVFP4 release is intended for efficient serving in runtimes with Gemma 4 NVFP4, reasoning, and tool-calling support.

Intended use

• OpenCode coding workflows
• Agentic software engineering tasks
• Long-context codebase navigation
• Reasoning-enabled tool use
• Multi-file implementation and editing sessions

Run with reasoning and tool calling enabled.

Format

• Format: NVFP4 / compressed-tensors
• Architecture: Gemma 4
• Total text parameters: ~48.1B
• Active text parameters: ~9.5B
• Context length: 262,144 tokens
• Language focus: English

Example outputs

In one continuous OpenCode chat, this checkpoint one-shot generated:

• a Flappy Bird-style game
• a Breakout game
• a skiing game
• a monster truck jumping over cars

Serving

Use a Gemma 4 compatible runtime with reasoning and tool calling enabled. For NVFP4 serving, use a runtime build that supports Gemma 4 compressed-tensors NVFP4 MoE kernels.

Example:

vllm serve <repo-or-local-path> \
  --trust-remote-code \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --enable-auto-tool-choice

For best results in OpenCode, keep the model's native chat template and enable thinking/reasoning mode.