gemma4-52b-a6b

Gemma4-52B-A6B

gemma4-52b-a6b is a production MoE expansion of unsloth/Gemma-4-26B-A4B-it. It was built with a two-GPU local training setup and expanded with additional specialist capacity for software engineering, repository reasoning, agentic workflows, and general instruction following.

This is a full model checkpoint, not a LoRA adapter.

Model Details

• Base model: unsloth/Gemma-4-26B-A4B-it
• Architecture: Gemma4 MoE
• Total logical parameters: approximately 50.6B
• Active parameters: approximately 6B
• Expert layout: 128 base experts + 128 added specialist experts
• Native context length: 256k tokens
• Validated local serve window: 131k tokens
• Primary focus: SWE, code/repository analysis, agentic workflows, and reasoning
• Recommended temperature: 0.5 to 0.7

Serving

Use the included Gemma4 chat template. Thinking mode is intended to be enabled.

CUDA_VISIBLE_DEVICES=0 \
vllm serve . \
  --served-model-name gemma4-52b-a6b \
  --host 0.0.0.0 \
  --port 23333 \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.88 \
  --trust-remote-code \
  --reasoning-parser gemma4 \
  --tool-call-parser gemma4 \
  --enable-auto-tool-choice \
  --chat-template ./chat_template.jinja \
  --default-chat-template-kwargs '{"enable_thinking": true}'

OpenAI-compatible endpoint:

http://localhost:23333/v1

Notes

This release is intended for production evaluation of the expanded Gemma4 MoE line. Feedback is welcome through the model discussion page or the usual Hugging Face repository feedback channels.