Gemma 4 (12B, Q8)
Google's Gemma 4 at 12B parameters, 8-bit quantization for higher fidelity than the Q4 build (at roughly double the memory footprint). Encoder-free unified multimodal model (text + image + audio + video), native tool calling, configurable thinking, 256K context. Apache-2.0 licensed.
At a glance
- Parameters
- 12B
- Quantization
- Q8_0
- Context window
- 256K tokens
- Approx. size
- 12.7 GB
- Engines
- llama.cpp, MLX
- License
- Apache-2.0 (open)
- Version
- v1.0.0 , released 2026-06-03
- Family
- gemma
Tuning
Defaults a gezel client applies out of the box, from real evaluation runs against this exact quantization.
- sampling.temperature
1- sampling.topP
0.95- sampling.topK
64- sampling.maxTokens
8192- sampling.repetitionPenalty
1.1- sampling.repetitionContext
20- reasoning.thinkingBudget
256- reasoning.enableThinking
true- profiles
- thinking-general, thinking-coding, thinking-precise, instruct, creative
Model behaviors
Client-side behavior modules the catalog enables for this model (reasoning-tag handling, fabrication detection, prompt shaping).
- reasoning.strip-channel-tags
- reasoning.capture-pre-tool-prose
- prompt.private-reasoning-guidance
- prompt.tool-cookbook-full
- prompt.meester-build-prelude
- fabrication.detect-past-tense-no-tools
- fabrication.detect-claim-without-tool
- turn.preamble-folding
- turn.ramble-detection
- turn.auto-acknowledge-tool-errors
- turn.permission-stall
- turn.continuation-budget
- turn.single-tool-per-turn
- mcp.compact-tool-schemas
- mcp.relax-required-fields
- mcp.default-missing-fields
- mcp.validate-ids-strict
- parse.gemma-special-token
- prompt.prefer-writefile-edits
- prompt.derive-by-execution
- prompt.terse-visible-reply
- tools.mlx-grammar
Sources
- llama.cpp (GGUF)
- ggml-org/gemma-4-12B-it-GGUF ·
gemma-4-12B-it-Q8_0.gguf· Q8_0 - MLX
- mlx-community/gemma-4-12b-it-8bit
- Upstream
- https://huggingface.co/google/gemma-4-12B-it
- Manifest
- View on GitHub
Tags
- multimodal
- vision
- audio
- tools
- agentic