satgeze/ gemma4-12b-uncensored-1.5m:latest

76 Downloads Updated 16 hours ago

Gemma 4 12B uncensored, 1.5M max context: certified 10/10 to 1.31M on a 32GB card. Vision included.

vision tools thinking

ollama run satgeze/gemma4-12b-uncensored-1.5m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/gemma4-12b-uncensored-1.5m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/gemma4-12b-uncensored-1.5m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/gemma4-12b-uncensored-1.5m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 16 hours ago

16 hours ago

4eabc5be7dfb · 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, 1.5M context

The beyond-1M build: same HauhauCS uncensored Gemma-4-12B trunk as gemma4-12b-uncensored-1m, baked with YaRN factor 6 for a 1,572,864-token maximum. The longest needle-certified context we know of, and the whole certification ran on a single 32GB RTX 5090. Native renderer and parser set:

ollama launch claude --model satgeze/gemma4-12b-uncensored-1.5m

Verified, not claimed

Tier	Context	Result
Certified (f16 KV)	262K / 1M / 1,310,720	¹⁰⁄₁₀ at every rung
Budget (q8_0 KV)	1M	¹⁰⁄₁₀
Budget (q8_0 KV)	1,572,864	⁹⁄₁₀, the miss is published

We also tested a factor-8 build to 2M: it scored ⁶⁄₁₀ there, below our bar, so it stays unshipped and the full study is published instead (beyond-1M study). NVIDIA RULER results are next for this build.

Tags

Tag	Size	Note
`latest`	7.6GB	The only quant: Google ships this QAT checkpoint in 4-bit only. Vision bundled

Raise context with /set parameter num_ctx (up to 1572864; Gemma’s sliding-window attention keeps KV tiny). For the certified experience stay at or below 1310720.

MTP note

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

Links

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

Credits: Google (Gemma 4 + QAT), HauhauCS (uncensoring), Unsloth (MTP head), SatGeze (1.5M extension, beyond-1M certification, 2M ceiling study).

![Gemma 4 1.5M](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1.5M-GGUF/resolve/main/banner.jpeg)

# Gemma 4 12B Uncensored, 1.5M context

The beyond-1M build: same HauhauCS uncensored Gemma-4-12B trunk as gemma4-12b-uncensored-1m, baked with YaRN factor 6 for a 1,572,864-token maximum. The longest needle-certified context we know of, and the whole certification ran on a single 32GB RTX 5090. Native renderer and parser set:

```
ollama launch claude --model satgeze/gemma4-12b-uncensored-1.5m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1.5M-GGUF/resolve/main/niah_heatmap.png)

| Tier | Context | Result |
|---|---|---|
| Certified (f16 KV) | 262K / 1M / **1,310,720** | **10/10 at every rung** |
| Budget (q8_0 KV) | 1M | 10/10 |
| Budget (q8_0 KV) | 1,572,864 | 9/10, the miss is published |

We also tested a factor-8 build to 2M: it scored 6/10 there, below our bar, so it stays unshipped and the full study is published instead ([beyond-1M study](https://github.com/satindergrewal/aviary-1m/blob/master/docs/beyond-1m-study.md)). NVIDIA RULER results are next for this build.

## Tags

| Tag | Size | Note |
|---|---|---|
| `latest` | 7.6GB | The only quant: Google ships this QAT checkpoint in 4-bit only. Vision bundled |

Raise context with `/set parameter num_ctx` (up to 1572864; Gemma's sliding-window attention keeps KV tiny). For the certified experience stay at or below 1310720.

## MTP note

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

## Links

Full repo, study data, heatmap: [Hugging Face](https://huggingface.co/satgeze/Gemma4-12B-Uncensored-HauhauCS-1.5M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Gemma4-12B-Uncensored-HauhauCS-1.5M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: Google (Gemma 4 + QAT), HauhauCS (uncensoring), Unsloth (MTP head), SatGeze (1.5M extension, beyond-1M certification, 2M ceiling study).

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