Gemma 4 12B

EXL3 3 bpw · tabbyapi · 32K context · RTX 5060 · 8 GB

Google's Gemma 4 12B from May 2026: a compact multimodal model with thinking, good with images and documents.

36 tok/sdecode

755 tok/sprefill

32Kcontext window

Tested on this cardSep 27, 2026

Run it

Follow the steps below to download and run this recipe. Omarchy Local AI lists the recipes it can launch on your hardware. The server then answers on http://localhost:8000/v1.

1 Download the weights

hf download turboderp/gemma-4-12B-it-exl3 \
  --revision 38309570753c5fde71a818feb513b1e09a81fc18 \
  --local-dir ~/models/gemma-4-12B-it-exl3-38309570

2 Write the server config

cat > gemma-4-12b.tabbyapi.32k.yml <<'EOF'
network: {host: 0.0.0.0, port: 5000, disable_auth: true, disable_fetch_requests: true, send_tracebacks: false, api_servers: [OAI], sse_ping_interval: 15}
logging: {log_prompt: false, log_generation_params: false, log_requests: false, log_chat_completion_requests: false}
model:
  model_dir: /workspace/models
  model_name: gemma-4-12B-it-exl3-38309570
  backend: exllamav3
  inline_model_loading: false
  max_seq_len: 32768
  cache_size: 33792
  cache_mode: Q4
  tensor_parallel: false
  gpu_split_auto: true
  autosplit_reserve: [256]
  chunk_size: 2048
  output_chunking: true
  max_batch_size: 1
  vision: true
  reasoning: true
  start_in_reasoning: auto
  tool_format: null
  template_vars_default: {enable_thinking: true}
draft_model: {}
memory: {sysmem_recurrent_cache: 4096, sysmem_kv_cache: 0, cuda_malloc_async: true}
EOF

3 Start the server

docker run --rm \
  --gpus all \
  -p 8000:5000 \
  --shm-size 8g \
  -v ~/models/gemma-4-12B-it-exl3-38309570:/workspace/models/gemma-4-12B-it-exl3-38309570:ro \
  -v $PWD/gemma-4-12b.tabbyapi.32k.yml:/app/config.yml:ro \
  --entrypoint /opt/venv/bin/python3 \
  ghcr.io/0xsero/tabbyapi-exl3@sha256:0f83e6198dc3be2561652d8df2525a7d1a69733e12fc8e25cbc5bc319a8a3ad1 \
  main.py \
  --config /app/config.yml

What it passed

✓ loadLoads and serves within an hour

✓ chatAnswers a plain question and stops on its own

✓ reasoningThinks separately and gets 17 × 23 right

✓ toolsCalls a tool with the right arguments and uses the result

✓ contextRecalls a code buried in a prompt that fills 85% of the window

✓ speedDecodes at 15 tok/s or more

Passed all six checks on a real RTX 5060 on Sep 27, 2026.

Details

Weights: turboderp/gemma-4-12B-it-exl3 @ 3830957075 ›
Image: ghcr.io/0xsero/tabbyapi-exl3@sha256:0f83e6198dc3be2561652d8df2525a7d1a69733e12fc8e25cbc5bc319a8a3ad1
Engine profile: tabbyapi-exl3 ›
Recipe file: registry/recipes/nvidia/rtx-5060-8gb/gemma-4-12b.tabbyapi.32k.json ›
Model released: May 23, 2026
Tested: vast, RTX 5060