1M CONTEXTMTPVISIONUNCENSORED

QwenPaw-Flash-9B-heretic-1M

Agent-optimized 9B, uncensored, with a needle-verified million-token context

SC117's heretic build of QwenPaw-Flash-9B (agent-trajectory finetune of Qwen3.5-9B) with YaRN rope scaling baked into the GGUF metadata for a 1,048,576-token context window, 4x the native 262,144. Weights are bit-identical to SC117's release, which already carries the official Qwen3.5-9B MTP layer injected back after the QwenPaw fine-tuning stripped it.

Uncensored (Heretic v1.3.0) · MTP baked in · Vision tower included · 1M needle-verified

Verified on these exact files

Capability	Result
1M context	10 needles per rung, depths 5 to 95 percent, temp 0, Q8_0 + f16 KV: 10/10 at every rung from 64K through 524K on an RTX 5090; 786K and 1M rungs running on a 128 GB M3 Max, card updates when they land
MTP speculative decoding	217.9 to 273.2 tok/s (+25 percent), draft acceptance 0.702, output identical by construction
Vision	mmproj tower reads image text and identifies objects correctly
Coherence	Q8_0 and Q4_K_M both pass the repetition-collapse gate

Raw per-needle records including every run: results.jsonl.

Files

File	Size	Pick it when
`qwenpaw-9b-1M-MTP-Q8_0.gguf`	9.8 GB	Max quality. On a 128 GB Mac this runs the full 1M with f16 KV (~43 GB total)
`qwenpaw-9b-1M-MTP-Q4_K_M.gguf`	5.8 GB	32 GB GPUs. Full 1M fits with q8_0 KV (budget config); ~524K at f16 KV
`mmproj-qwenpaw-9b.gguf`	0.9 GB	Vision, attach with `--mmproj`

No other quants on purpose: the 9B is small enough that Q8_0 is the sensible default and Q4_K_M covers the budget case.

Every file, every mirror

Nothing was discontinued: every quant is one click away. Hugging Face carries the curated picks, ModelScope always carries everything, and Ollama serves ready-to-run tags.

On Ollama every tag ships with the vision tower bundled and the 1M rope metadata baked in.

File	Size	Hugging Face	ModelScope	Ollama
`mmproj-qwenpaw-9b.gguf`	922 MB	download	download	bundled in every tag
`qwenpaw-9b-1M-MTP-Q4_K_M.gguf`	5.8 GB	download	download	`ollama run satgeze/qwenpaw-9b-heretic-1m:q4_k_m`
`qwenpaw-9b-1M-MTP-Q8_0.gguf`	9.8 GB	download	download	`ollama run satgeze/qwenpaw-9b-heretic-1m`

Run it

llama-server -m qwenpaw-9b-1M-MTP-Q8_0.gguf \
  -c 1048576 -np 1 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --mmproj mmproj-qwenpaw-9b.gguf

Ollama (1M and vision work; no speculative decoding in Ollama yet):

FROM ./qwenpaw-9b-1M-MTP-Q8_0.gguf
RENDERER qwen3.5
PARSER qwen3.5
PARAMETER num_ctx 262144

How this was built

YaRN rope-scaling metadata (factor 4.0 over native 262,144) written into the GGUF header with gguf-py. No weight changes, no fine-tuning by us. Certification: multi-needle harness against llama-server, f16 KV only for cert runs. Method and tooling: github.com/satindergrewal/aviary-1m.

Note on upstream speed claims: SC117 reports up to 4.1x on time-scored agent benchmarks. Our controlled A/B on identical prompts measures +25 percent decode; speculative decoding cannot change model outputs at temperature 0, so treat benchmark-score deltas from MTP as timing artifacts.

Credits

Base: Qwen (Apache-2.0), including the MTP layer and vision tower. Agent fine-tune: agentscope-ai. Heretic abliteration and MTP re-injection: SC117. 1M YaRN extension and certification: SatGeze.

Mirrors: Hugging Face | ModelScope. Sister repos: Uncensored 1M collection