satgeze/ qwen36-35b-uncensored-1m

72 Downloads Updated 20 hours ago

Qwen3.6-35B uncensored flagship: 1M certified 70/70, vision, MTP grafted. One file, four capabilities.

vision tools thinking

ollama run satgeze/qwen36-35b-uncensored-1m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/qwen36-35b-uncensored-1m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/qwen36-35b-uncensored-1m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/qwen36-35b-uncensored-1m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code

Claude Code ollama launch claude --model satgeze/qwen36-35b-uncensored-1m

Codex App

Codex App ollama launch codex-app --model satgeze/qwen36-35b-uncensored-1m

OpenClaw

OpenClaw ollama launch openclaw --model satgeze/qwen36-35b-uncensored-1m

Hermes Agent

Hermes Agent ollama launch hermes --model satgeze/qwen36-35b-uncensored-1m

Codex

Codex ollama launch codex --model satgeze/qwen36-35b-uncensored-1m

OpenCode

OpenCode ollama launch opencode --model satgeze/qwen36-35b-uncensored-1m

Models

Name

2 models

Size / Usage

Context

Input

qwen36-35b-uncensored-1m:latest

23GB · 1M context window · Text, Image · 21 hours ago

qwen36-35b-uncensored-1m:latest

23GB

1M

Text, Image

qwen36-35b-uncensored-1m:q4_k_m-no-mtp

22GB · 1M context window · Text, Image · 20 hours ago

qwen36-35b-uncensored-1m:q4_k_m-no-mtp

22GB

1M

Text, Image

Readme

Qwen3.6-35B Uncensored, 1M certified

HauhauCS’s Aggressive uncensored Qwen3.6-35B-A3B with a 1,048,576-token context baked in, the official MTP layer grafted back at the tensor level, and vision attached. One file, four capabilities. Native renderer and parser set:

ollama launch claude --model satgeze/qwen36-35b-uncensored-1m

Verified, not claimed

A perfect ⁷⁰⁄₇₀: 10 needles at every rung from 64K through 1M, f16 KV, temperature 0. Not a single miss anywhere in the ladder.

Pick your tag

Tag	Size	When
`latest`	23GB	The full package: uncensored + 1M + MTP + vision
`q4_k_m-no-mtp`	22GB	Plain uncensored 1M trunk without the MTP layer

Hybrid GDN architecture keeps KV small: a 128GB Mac runs the full 1M in about 44GB total; a 32GB card holds roughly 262K fully resident. Raise with /set parameter num_ctx.

MTP note

The grafted MTP layer measured +34 percent decode with acceptance 0.864 under llama.cpp; dormant in Ollama until speculative decoding ships there, then it activates automatically.

Links

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

Credits: Qwen (base + MTP layer, Apache-2.0), HauhauCS (uncensoring), Unsloth (MTP GGUF packaging), SatGeze (1M extension, MTP graft, certification).

![Qwen3.6](https://huggingface.co/satgeze/Qwen3.6-35B-Uncensored-HauhauCS-1M-GGUF/resolve/main/banner.jpeg)

# Qwen3.6-35B Uncensored, 1M certified

HauhauCS's Aggressive uncensored Qwen3.6-35B-A3B with a 1,048,576-token context baked in, the official MTP layer grafted back at the tensor level, and vision attached. One file, four capabilities. Native renderer and parser set:

```
ollama launch claude --model satgeze/qwen36-35b-uncensored-1m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/Qwen3.6-35B-Uncensored-HauhauCS-1M-GGUF/resolve/main/qwen36_heatmap.png)

**A perfect 70/70**: 10 needles at every rung from 64K through 1M, f16 KV, temperature 0. Not a single miss anywhere in the ladder.

![MTP speedup](https://huggingface.co/satgeze/Qwen3.6-35B-Uncensored-HauhauCS-1M-GGUF/resolve/main/qwen36_mtp_speedup.png)

## Pick your tag

| Tag | Size | When |
|---|---|---|
| `latest` | 23GB | The full package: uncensored + 1M + MTP + vision |
| `q4_k_m-no-mtp` | 22GB | Plain uncensored 1M trunk without the MTP layer |

Hybrid GDN architecture keeps KV small: a 128GB Mac runs the full 1M in about 44GB total; a 32GB card holds roughly 262K fully resident. Raise with `/set parameter num_ctx`.

## MTP note

The grafted MTP layer measured +34 percent decode with acceptance 0.864 under llama.cpp; dormant in Ollama until speculative decoding ships there, then it activates automatically.

## Links

Full repo, raw data, charts: [Hugging Face](https://huggingface.co/satgeze/Qwen3.6-35B-Uncensored-HauhauCS-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Qwen3.6-35B-Uncensored-HauhauCS-1M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: Qwen (base + MTP layer, Apache-2.0), HauhauCS (uncensoring), Unsloth (MTP GGUF packaging), SatGeze (1M extension, MTP graft, certification).

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