Released quietly, reused constantly

The exact artifact is 9,653,104,368 parameters of dense Qwen3.5-9B in BF16 — about 19.3 GB on disk — pinned at revision 05b9e7c9b978ba29bdb8f50a49c30e4b91183339. The Hugging Face model API shows 9,195 downloads and 125 likes as of this writing. Respectable for an uncensored 9B. Unremarkable next to the hype machines.

What is remarkable sits downstream. A search for the model name returns 58 other repositories built on it, with 64,688 combined downloads. mradermacher's GGUF conversion alone counts 34,479 downloads and 36 likes — about 3.7× the base's own download count. The base model is out-downloaded by its own derivatives. That is the signature of a quiet classic: it is the middle of the chain, not the destination.

A dense 9B with a hybrid spine

Qwen3.5-9B is dense, not sparse: 32 layers, hidden size 4096, arranged as eight repeats of three Gated DeltaNet blocks followed by one Gated Attention block, per the upstream card. The DeltaNet linear-attention layers are what let a 9B carry a 262,144-token native context without the KV-cache blowup of pure full attention. It also ships a vision encoder — the pipeline tag is image-text-to-text — so this refusal edit lands on a multimodal model. Abliteration changes weights, not architecture.

A proof of concept, published on purpose

The model card is unusually candid: “a crude, proof-of-concept implementation to remove refusals from an LLM model without using TransformerLens,” pointing at Sumandora's remove-refusals-with-transformers. No refusal-rate table. No post-edit eval. The practical advice is an Ollama one-liner — ollama run huihui_ai/qwen3.5-abliterated:9b — plus the standard warnings about reduced safety filtering and research-only use.

It was not the first Qwen3.5 abliteration: huihui-ai had already published 27B and 35B-A3B versions on 27 February, and the 27B sibling has since pulled 161,669 downloads — seventeen times this one. The 9B arrived 9 March, inside the family's first fortnight, and it is the one that became the default for small hardware.

The derivative economy

The 58 downstream repos split into two camps. Conversion: GGUF from mradermacher, nuofang, Abiray and others; AWQ, GPTQ and exl3 from nicklas373, nemozxy123 and groxaxo; NVFP4 from sakamakismile; an official MLX 4-bit from huihui-ai itself plus vanch007's MLX line. Coverage for every runtime people actually own.

The second camp is the telling one: people trained on the abliterated weights. nbeerbower's Grimoire series — ORPO, SFT, DPO, KTO, SimPO, TIES, TIES-Hemlock-SFT — is a preference-tuning family whose cards declare the base is this exact repo. Abliteration was not the endpoint; it was the starting point for further edits. The sibling Huihui-Qwen3.5-9B-Claude-4.6-Opus-abliterated spun a smaller tree of its own. One clarification for our catalog: Qwythos shares the Qwen3.5-9B lineage but builds directly on Qwen's checkpoint, not on this repo.

Word of mouth, not benchmarks

The community record is thin but consistent. The model's first discussion, opened the day it was published, is titled “Perfect size”: “It fits perfect in a RTX 3060.” A 12 GB card. That is the actual product: an Apache-2.0 multimodal model with real coding and reasoning chops that runs on hardware people already own, installable with one command.

The second thread is a working conversation about running the same recipe on other bases, where one participant calls huihui-ai's ablation models “the most effective” — a community opinion, cited as such. HackerNoon carried a roundup in April 2026 recommending it for controlled research use. It contains no benchmarks either. The popularity is a social fact, not a measurement.

The numbers that don't exist

huihui-ai publishes no post-edit evaluation for this checkpoint. The upstream Qwen3.5-9B card reports MMLU-Pro 82.5, GPQA Diamond 81.7, IFEval 91.5, LiveCodeBench v6 65.6, BFCL-V4 66.1 and AA-LCR 63.0. Strong numbers, and they belong to Qwen's checkpoint, not this one. Nothing in this repo proves the edit preserved them.

How we treat it

We pin revision 05b9e7c9b978ba29bdb8f50a49c30e4b91183339, serve a conservative context and budget a single L40S; the approximate managed price estimate is $2.34/hour. We do not transfer upstream benchmarks to our endpoint until the exact artifact is measured. “Abliterated” still means refusal-reduced — not zero-refusal, not harmless, not legal everywhere.

The idea, in plain words

Why one model becomes 58 repositories

Same brain, different suitcases. GGUF, Q4_K_M, AWQ, MLX, FP8 — these are compression formats tuned for different hardware: llama.cpp, Apple Silicon, NVIDIA, AMD. Every one of the 58 derivative repos is the same Qwen3.5-9B under a different packing, which is why downloads stack up across the family instead of on the base.

Primary sources