satgeze/ gemma4-26b-uncensored-1m:latest

45 Downloads Updated 20 hours ago

Gemma 4 26B-A4B uncensored MoE: 1M context, vision, fast. ~91% recall, honestly documented.

vision tools thinking

ollama run satgeze/gemma4-26b-uncensored-1m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/gemma4-26b-uncensored-1m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/gemma4-26b-uncensored-1m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/gemma4-26b-uncensored-1m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 20 hours ago

20 hours ago

73687620b6a2 · 18GB ·

model

archgemma4

·

parameters25.2B

·

quantizationQ4_K_M

17GB

projector

archclip

·

parameters573M

·

quantizationBF16

1.2GB

params

{ "num_ctx": 262144 }

19B

Readme

Gemma 4 26B-A4B Uncensored, 1M context

HauhauCS’s uncensored Gemma-4-26B-A4B (MoE, 4B active, Google QAT) with a 1,048,576-token context baked in and vision attached. The fastest of the uncensored Gemma trio. Native renderer and parser set:

ollama launch claude --model satgeze/gemma4-26b-uncensored-1m

Verified, with full honesty

Needle recall averages ~91 percent across two full seed sets: roughly one random needle per rung, at all lengths including native range. This is a small flat abliteration tax, not a long-context failure (the official censored trunk scores ¹⁰⁄₁₀ under the same harness). If a rare retrieval miss is unacceptable for your workload, use satgeze/gemma4-12b-uncensored-1m, which is certified clean. Every run including the misses is published in the repo data.

Tags

Tag	Size	Note
`latest`	18GB	The only quant in existence: Google ships this QAT checkpoint in 4-bit only

KV is about 18KB per token; raise context with /set parameter num_ctx.

MTP note

A separate 252MB draft head lives on the repo (+48 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, testing).

![Gemma 4](https://huggingface.co/satgeze/Gemma4-26B-A4B-Uncensored-HauhauCS-1M-GGUF/resolve/main/banner.jpeg)

# Gemma 4 26B-A4B Uncensored, 1M context

HauhauCS's uncensored Gemma-4-26B-A4B (MoE, 4B active, Google QAT) with a 1,048,576-token context baked in and vision attached. The fastest of the uncensored Gemma trio. Native renderer and parser set:

```
ollama launch claude --model satgeze/gemma4-26b-uncensored-1m
```

## Verified, with full honesty

![NIAH heatmap](https://huggingface.co/satgeze/Gemma4-26B-A4B-Uncensored-HauhauCS-1M-GGUF/resolve/main/niah_heatmap.png)

Needle recall averages **~91 percent** across two full seed sets: roughly one random needle per rung, at all lengths including native range. This is a small flat abliteration tax, not a long-context failure (the official censored trunk scores 10/10 under the same harness). If a rare retrieval miss is unacceptable for your workload, use `satgeze/gemma4-12b-uncensored-1m`, which is certified clean. Every run including the misses is published in the repo data.

## Tags

| Tag | Size | Note |
|---|---|---|
| `latest` | 18GB | The only quant in existence: Google ships this QAT checkpoint in 4-bit only |

KV is about 18KB per token; raise context with `/set parameter num_ctx`.

## MTP note

A separate 252MB draft head lives on the repo (+48 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/Gemma4-26B-A4B-Uncensored-HauhauCS-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Gemma4-26B-A4B-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, testing).

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