Gemma 4 (26B, MoE, Q4)
Google's Gemma 4 mixture-of-experts model — 25.2B total parameters with 3.8B active per token. Multimodal with long-context agentic workflows. The full 26B has to fit in memory but only ~4B parameters fire per token, so on capable hardware it runs noticeably faster than the dense 31B at a similar memory cost — the recommended high-end on-device default.
At a glance
- Parameters
- 25.2B
- Quantization
- UD-Q4_K_XL
- Context window
- 256K tokens
- Approx. size
- 14.2 GB
- Engines
- llama.cpp, MLX, Ollama
- License
- Apache-2.0 (open)
- Version
- v1.2.0 , released 2026-06-06
- 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
96- 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.continuation-budget
- mcp.validate-ids-strict
- turn.single-tool-per-turn
- parse.gemma-special-token
- tools.mlx-grammar
Sources
- llama.cpp (GGUF)
- unsloth/gemma-4-26B-A4B-it-qat-GGUF ·
gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf· UD-Q4_K_XL - MLX
- mlx-community/gemma-4-26B-A4B-it-qat-nvfp4
- Ollama
gemma4:26b-a4b-it-qat- Upstream
- https://huggingface.co/google/gemma-4-26B-A4B-it
- Manifest
- View on GitHub
Tags
- multimodal
- vision
- tools
- mix of experts
- agentic
- qat