Gemma 4 31B text-only instruct model with tool-calling and thinking support, quantized to Q3_K_S and optimized for 16 GB VRAM. Supports up to 256K context, with approximately 64K tokens fitting fully within 16 GB VRAM in tested configurations.

tools thinking 31b

ollama run byczech/gemma4-it-16G-Q3_K_S:31b

curl http://localhost:11434/api/chat \
  -d '{
    "model": "byczech/gemma4-it-16G-Q3_K_S:31b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='byczech/gemma4-it-16G-Q3_K_S:31b',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'byczech/gemma4-it-16G-Q3_K_S:31b',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 3 weeks ago

3 weeks ago

a3085fcdc8ec · 13GB ·

model

archgemma4

parameters30.7B

quantizationQ3_K_S

13GB

license

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US

11kB

params

{ "temperature": 1, "top_k": 64, "top_p": 0.95 }

42B

template

13B

Readme

Context length

The model supports a maximum context window of up to 256K tokens.

On a 16 GB GPU, approximately 64,000 tokens can fit fully within VRAM in tested configurations using a quantized KV cache. Slightly larger context sizes may begin to spill into system RAM, reducing inference performance.

The exact GPU-resident limit may vary depending on:

Ollama version
GPU architecture and driver
Compute backend
KV-cache type
Runtime overhead
Other applications using GPU memory
Parallel requests and loaded models

For a good balance between context size and fully GPU-resident inference on a 16 GB GPU, a context size of around 64K tokens is recommended.