gemma4:31b-mtp-vision-bf16

vision tools thinking

ollama run home-rag/gemma4:31b-mtp-vision-bf16

curl http://localhost:11434/api/chat \
  -d '{
    "model": "home-rag/gemma4:31b-mtp-vision-bf16",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='home-rag/gemma4:31b-mtp-vision-bf16',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'home-rag/gemma4:31b-mtp-vision-bf16',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 3 days ago

3 days ago

5f932aa60991 · 64GB ·

model

archgemma4

parameters30.7B

quantizationBF16

61GB

projector

archclip

parameters576M

quantizationBF16

1.2GB

draft

GGUF��1��-��general.architecture��gemma4-assistant��general.type

955MB

template

13B

params

{ "draft_num_predict": 2, "temperature": 0.95, "top_k": 64, "top_p": 0.95 }

67B

Readme

Gemma 4 31B

ARCH: GEMMA_V4 // PARAMS: 31B // PRECISION: BF16 // CONTEXT: 256K TOKENS

Hardware & Environment

Designed for local multi-GPU systems and small GPU grids.

Target hardware: AMD Instinct MI50, MI60, MI100 and MI210 series
Recommended memory: 96+ GB HBM / VRAM

Technical Specifications

Model: Gemma 4 31B
Origin 1: Hugging Face google/gemma-4-31B-it
Origin 2: Hugging Face google/gemma-4-31B-it-assistant
Precision: BF16
Vision: Integrated mmproj (archclip) for multimodal input
Acceleration: Multi-Token Prediction (MTP) draft head
Context: Up to 262144 tokens

Quick Start

ollama run home-rag/gemma4:31b-mtp-vision-bf16