---

base_model: google/gemma-4-31B-it

library_name: gguf

pipeline_tag: text-generation

license: other

license_name: saturn-community-license

license_link: LICENSE

tags:

- gemma4

- gguf

- q4_k_m

- code

- html

---

Saturn 31B — Q4_K_M GGUF

A fine-tune of Gemma 4 31B (dense, instruct), trained on Fireworks AI and

merged + quantized to Q4_K_M GGUF for local inference with llama.cpp.

Built for app and UI generation — HTML, frontend components, and small self-contained web applications.

License summary

You may use this model to build and ship applications, including commercial

ones. You may not resell or redistribute the model weights themselves as a

product, or present the model as your own. This model is a derivative of

Google Gemma 4 and remains subject to the Gemma Terms of Use, which apply in

addition to the terms below. See the LICENSE file for full text.

Specs

• Base: google/gemma-4-31B-it (text-only export)
• Quant: Q4_K_M (~4.87 bpw, ~17.4 GB)
• Context: up to 262,144 tokens

Recommended hardware

Total memory (VRAM + RAM) should exceed the file size (~17.4 GB).

• ~24 GB+ VRAM (RTX 3090/4090): full GPU offload, fast
• 16 GB VRAM: partial offload, usable
• CPU-only: works, but very slow

Run with llama.cpp

\\\`

llama-server -m saturn-31b-Q4_K_M.gguf -ngl 99 -c 32768 --jinja

\\\`

Gemma 4 has a thinking mode; disable with

\--chat-template-kwargs '{"enable_thinking":false}'\ for direct answers