satgeze/ qwenpaw-9b-heretic-1m:Q4_K_M

75 Downloads Updated 20 hours ago

Agent-optimized 9B heretic: 1M context, MTP, vision. 50/50 needles to 524K, uncensored.

vision tools thinking

ollama run satgeze/qwenpaw-9b-heretic-1m:Q4_K_M

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/qwenpaw-9b-heretic-1m:Q4_K_M",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/qwenpaw-9b-heretic-1m:Q4_K_M',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/qwenpaw-9b-heretic-1m:Q4_K_M',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 20 hours ago

20 hours ago

49188287281b · 6.7GB ·

model

archqwen35

·

parameters9.2B

·

quantizationQ4_K_M

5.8GB

projector

archclip

·

parameters456M

·

quantizationBF16

922MB

params

{ "num_ctx": 262144 }

19B

Readme

QwenPaw-Flash-9B heretic, 1M context

SC117’s heretic-abliterated build of agentscope-ai’s QwenPaw-Flash-9B (an agent-optimized finetune of Qwen3.5-9B), with a 1,048,576-token context baked in, the official MTP layer, and vision attached. Uncensored, agent-tuned, needle-tested. Native renderer and parser set, so tool calling works out of the box:

ollama launch claude --model satgeze/qwenpaw-9b-heretic-1m

Verified, not claimed

⁵⁰⁄₅₀ needles at every rung from 64K through 524K (Q8_0, f16 KV, temperature 0, RTX 5090). The 786K and 1M rungs are running on a 128GB Mac at publish time; this page and the heatmap update when they land.

Tags

Tag	Size	When
`latest`	9.8GB	Q8_0, max quality, MTP layer baked in
`q4_k_m`	5.8GB	Smaller, 32GB-friendly

Both carry the 1M rope metadata and the vision tower. Raise context with /set parameter num_ctx.

MTP note

Measured +25 percent decode under llama.cpp with speculation active (acceptance 0.70). Dormant in Ollama until speculative decoding ships there, then it activates automatically with no re-pull. On the upstream “1.7 to 4.1x” claims: those were time-scored agent benchmarks; speculative decoding cannot change model outputs at temperature 0, so treat score gains from MTP as timing artifacts. The honest, controlled decode gain is +25 percent.

Links

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

Credits: Qwen (base + MTP layer, Apache-2.0), agentscope-ai (QwenPaw agent finetune), SC117 (heretic abliteration + MTP re-injection), SatGeze (1M extension, certification).

![QwenPaw](https://huggingface.co/satgeze/QwenPaw-Flash-9B-heretic-1M-GGUF/resolve/main/banner.jpeg)

# QwenPaw-Flash-9B heretic, 1M context

SC117's heretic-abliterated build of agentscope-ai's QwenPaw-Flash-9B (an agent-optimized finetune of Qwen3.5-9B), with a 1,048,576-token context baked in, the official MTP layer, and vision attached. Uncensored, agent-tuned, needle-tested. Native renderer and parser set, so tool calling works out of the box:

```
ollama launch claude --model satgeze/qwenpaw-9b-heretic-1m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/QwenPaw-Flash-9B-heretic-1M-GGUF/resolve/main/niah_heatmap.png)

50/50 needles at every rung from 64K through 524K (Q8_0, f16 KV, temperature 0, RTX 5090). The 786K and 1M rungs are running on a 128GB Mac at publish time; this page and the heatmap update when they land.

![MTP speedup](https://huggingface.co/satgeze/QwenPaw-Flash-9B-heretic-1M-GGUF/resolve/main/mtp_speedup.png)

## Tags

| Tag | Size | When |
|---|---|---|
| `latest` | 9.8GB | Q8_0, max quality, MTP layer baked in |
| `q4_k_m` | 5.8GB | Smaller, 32GB-friendly |

Both carry the 1M rope metadata and the vision tower. Raise context with `/set parameter num_ctx`.

## MTP note

Measured +25 percent decode under llama.cpp with speculation active (acceptance 0.70). Dormant in Ollama until speculative decoding ships there, then it activates automatically with no re-pull. On the upstream "1.7 to 4.1x" claims: those were time-scored agent benchmarks; speculative decoding cannot change model outputs at temperature 0, so treat score gains from MTP as timing artifacts. The honest, controlled decode gain is +25 percent.

## Links

Full repo, raw data, heatmap: [Hugging Face](https://huggingface.co/satgeze/QwenPaw-Flash-9B-heretic-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/QwenPaw-Flash-9B-heretic-1M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: Qwen (base + MTP layer, Apache-2.0), agentscope-ai (QwenPaw agent finetune), SC117 (heretic abliteration + MTP re-injection), SatGeze (1M extension, certification).

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