AtomicChat/gemma-4-31B-it-GGUF

Base model: google/gemma-4-31B-it

Gemma 4 31B, self-quantized to GGUF by Atomic Chat. Built straight from Google's original weights with a per-tensor importance matrix. Runs fully offline.

Highlights

• Multimodal — natively handles text and image input and generates text output.
• Built-in reasoning — designed as a capable reasoner with a configurable thinking mode set via the system prompt.
• 256K context window for long documents and codebases.
• Multilingual — out-of-the-box support for 35+ languages, pre-trained on 140+ languages.
• Coding & agentic — native function-calling support and notable coding-benchmark improvements.
• Native system prompt — native support for the system role for more structured conversations.

These GGUFs are self-quantized from the original weights, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.

Always pass --jinja so the Gemma 4 31B chat template is applied. Without it the model can emit malformed turns.

Model Overview

Property	Value
Base model	google/gemma-4-31B-it
Total parameters	30.7B (Dense)
Layers	60
Context length	256K tokens
Vocabulary	262K
Architecture	Dense decoder, hybrid local/global attention with Proportional RoPE
This repo	GGUF quants (imatrix) + vision mmproj

Gemma 4 31B is multimodal. This repo ships the mmproj-gemma4-31b-it-f16.gguf vision projector. With -hf it is pulled automatically; otherwise pass --mmproj. Use llama-mtmd-cli or llama-server to feed images.

Scores are Google's published results for the base google/gemma-4-31B-it. Quantization preserves the large majority of this; Q4_K_M and up sit within a point or two of full precision.

Choosing a quant

Quant	Size	Notes
Q2_K	11.9 GB	Smallest. Minimal RAM, clear quality drop.
IQ3_M	14.2 GB	Beats Q3 at similar size thanks to imatrix. Best low-RAM pick.
Q3_K_M	15.3 GB	Low quality but usable.
Q3_K_L	16.6 GB	A step above Q3_K_M.
IQ4_XS	16.7 GB	Excellent quality for size. Recommended low-bit.
Q4_K_S	17.8 GB	Compact Q4, fast.
Q4_K_M	18.7 GB	Recommended default. Best balance of size, speed and quality.
UD-Q4_K_XL	19.0 GB	Dynamic. Embeddings and output kept at Q8_0 for higher quality at a Q4 footprint.
Q5_K_S	19.6 GB	Higher quality.
Q5_K_M	17.4 GB	Higher quality, low loss.
Q6_K	7.6 GB	Near lossless.
Q8_0	18.4 GB	Effectively lossless, reference quality.

Pick the largest file that fits your (V)RAM with room for context. Q4_K_M or UD-Q4_K_XL is the sweet spot for most setups; Q6_K or Q8_0 for maximum fidelity.

Get started

Run Gemma 4 31B locally with:

• Atomic Chat: the easiest path. Open the app, search AtomicChat/gemma4-31b-it-GGUF, pick a quant, hit Use this model.
• llama.cpp: llama-server -hf AtomicChat/gemma4-31b-it-GGUF:Q4_K_M --jinja -c 8192
• Ollama: ollama run hf.co/AtomicChat/gemma4-31b-it-GGUF:Q4_K_M
• LM Studio / Jan: search the repo id, download any quant.

Best practices

Parameter	Value
temperature	1.0
top_p	0.95
top_k	64

Google's standardized sampling configuration recommended across all use cases.

Run in llama.cpp

git clone https://github.com/ggerganov/llama.cpp
cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server

./llama.cpp/build/bin/llama-server \
    -hf AtomicChat/gemma4-31b-it-GGUF:UD-Q4_K_XL \
    --jinja -ngl 99 -c 8192 -fa on

How these were made

1. Download google/gemma-4-31B-it (original weights).
2. Convert to f16 GGUF with llama.cpp.
3. Build an importance matrix over calibration_datav3 (100 chunks).
4. Quantize the full ladder with --imatrix.
5. UD-Q4_K_XL additionally pins the token-embedding and output tensors to Q8_0.

License

Original model by Google DeepMind, released under the Apache 2.0 license. Quantized by Atomic Chat.