satgeze/ gemma4-12b-uncensored-1m:latest

80 Downloads Updated 22 hours ago

Gemma 4 12B uncensored: first Gemma certified 10/10 at 1M, fits a 32GB card. Vision included.

vision tools thinking

ollama run satgeze/gemma4-12b-uncensored-1m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/gemma4-12b-uncensored-1m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/gemma4-12b-uncensored-1m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/gemma4-12b-uncensored-1m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 22 hours ago

22 hours ago

4d017656f27b · 7.6GB ·

model

archgemma4

·

parameters11.9B

·

quantizationQ4_K_M

7.4GB

projector

archclip

·

parameters52.4M

·

quantizationBF16

175MB

params

{ "num_ctx": 262144 }

19B

Readme

Gemma 4 12B Uncensored, 1M certified

HauhauCS’s uncensored Gemma-4-12B (Google QAT checkpoint) with a 1,048,576-token context baked in and vision attached. The first Gemma 4 we know of certified needle-perfect at the full million tokens, and the entire certification ran inside a 32GB GPU. Native renderer and parser set:

ollama launch claude --model satgeze/gemma4-12b-uncensored-1m

Verified, not claimed

¹⁰⁄₁₀ at every rung from 64K through 1M (f16 KV, temperature 0). One cell missed on a first seed and passed on a second; both runs are published in the repo data.

Tags

Tag	Size	Note
`latest`	7.6GB	The only quant in existence: Google ships this QAT checkpoint in 4-bit only, so this is the complete set

Gemma’s 5:1 sliding-window attention keeps KV tiny: the full 1M fits on a 32GB card. Raise with /set parameter num_ctx 1048576.

MTP note

A separate 254MB draft head lives on the repo (+51 percent decode under llama.cpp with -md); Ollama has no speculative decoding yet.

Links

Full repo, raw data, charts: Hugging Face | ModelScope Method and harness: github.com/satindergrewal/aviary-1m

Credits: Google (Gemma 4 + QAT; Gemma license), HauhauCS (uncensoring), Unsloth (MTP head), SatGeze (1M extension, certification).

![Gemma 4](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1M-GGUF/resolve/main/banner.jpeg)

# Gemma 4 12B Uncensored, 1M certified

HauhauCS's uncensored Gemma-4-12B (Google QAT checkpoint) with a 1,048,576-token context baked in and vision attached. The first Gemma 4 we know of certified needle-perfect at the full million tokens, and the entire certification ran inside a 32GB GPU. Native renderer and parser set:

```
ollama launch claude --model satgeze/gemma4-12b-uncensored-1m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1M-GGUF/resolve/main/niah_heatmap.png)

10/10 at every rung from 64K through 1M (f16 KV, temperature 0). One cell missed on a first seed and passed on a second; both runs are published in the repo data.

## Tags

| Tag | Size | Note |
|---|---|---|
| `latest` | 7.6GB | The only quant in existence: Google ships this QAT checkpoint in 4-bit only, so this is the complete set |

Gemma's 5:1 sliding-window attention keeps KV tiny: the full 1M fits on a 32GB card. Raise with `/set parameter num_ctx 1048576`.

## MTP note

A separate 254MB draft head lives on the repo (+51 percent decode under llama.cpp with `-md`); Ollama has no speculative decoding yet.

## Links

Full repo, raw data, charts: [Hugging Face](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Gemma4-12B-Uncensored-HauhauCS-1M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: Google (Gemma 4 + QAT; Gemma license), HauhauCS (uncensoring), Unsloth (MTP head), SatGeze (1M extension, certification).

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)