| --- |
| license: apache-2.0 |
| license_link: https://ai.google.dev/gemma/docs/gemma_4_license |
| base_model: google/gemma-4-26B-A4B-it |
| base_model_relation: finetune |
| language: |
| - en |
| pipeline_tag: image-text-to-text |
| library_name: vllm |
| inference: false |
| tags: |
| - gemma4 |
| - opencode |
| - coding-agent |
| - tool-use |
| - reasoning |
| - nvfp4 |
| - code |
| - agent |
| --- |
| |
| # Gemma 4 OpenCoder 40B-A8B |
|
|
| Gemma 4 OpenCoder is a continued Gemma 4 reasoning model focused on OpenCode-style coding workflows, tool use, and production software tasks. |
|
|
| It is intended for agentic coding environments where the model can reason, inspect context, call tools, edit files, and continue through multi-step implementation tasks. |
|
|
| The checkpoint adds roughly 14B parameters to the Gemma 4 26B-A4B base and was post-trained on approximately 30k OpenCode-oriented coding samples. |
|
|
| ## Model profile |
|
|
| | Field | Value | |
| |---|---:| |
| | Total parameters | ~40.1B | |
| | Active parameters | ~8.0B | |
| | Added parameters | ~14.3B | |
| | Context length | up to 256K tokens | |
| | Format | NVFP4 | |
| | Base model | `google/gemma-4-26B-A4B-it` | |
| | Primary focus | coding-agent workflows, reasoning, and tool calling | |
|
|
| ## Recommended use |
|
|
| Run this model in OpenCode or another compatible agent harness with reasoning and tool calling enabled. |
|
|
| Use a recent vLLM runtime with Gemma 4 NVFP4 support: |
|
|
| ```bash |
| vllm serve <repo-id> \ |
| --max-model-len 262144 |
| ``` |
|
|
| For best results, use the native Gemma 4 chat template and keep reasoning/tool-calling behavior enabled in the client stack. |
|
|
| ## Example OpenCode session |
|
|
| In one continuous OpenCode chat, the model one-shot generated: |
|
|
| - a stylish SVG Flappy Bird-style game |
| - a Breakout-style game |
| - a skiing game |
| - a monster truck jumping over cars |
|
|
| ## Screenshots |
|
|
|  |
|
|
|  |
|
|
|  |
|
|
|  |
|
|
| ## Notes |
|
|
| This checkpoint is intended for coding-agent workflows. It is not positioned as a general chat-only model. Use a harness that supports long context, reasoning, and structured tool use. |
|
|
