Four hops down a family tree in one morning

The artifact pinned above is one step short of a root. Its own model card names two bases: tommytracx/Thox1-27b and prism-ml/Bonsai-27B-gguf. Follow the chain and you get a lineage that reads like a relay:

  • Prism ML built Bonsai-27B, a Qwen3.6-27B derivative whose language weights are stored as a single sign bit each, plus a shared scale per group of 128. The full-size repo is gated; the GGUF and MLX packs are public and huge.
  • s3nh decensored it on 14 July with Heretic v1.4.0, working from the FP16 unpack (the gated repo's full-precision form): refusals went from 81/100 on the original to 6/100, with KL divergence 0.0033 against the original. Those are the editor's own measurements on the FP16 edit.
  • tommytracx quantized the decensor to a Q1_0 GGUF on 1 August as Thox1-27b, a THOX.ai release.
  • guell00 repackaged it this morning as Velum-Unbound-Uncensored: the same Q1_0 weights plus a dspark drafter file, a fresh card, a Brazil flag, and an MTP story.

Nothing in that chain is gated. The final artifact is two files: VELUM-UNBOUND-Q1_0.gguf at 4,667,606,336 bytes and dspark/dspark-VELUM-UNBOUND-Q1_0.gguf at 1,946,393,568 bytes, per the HF model API. Pinned revision 97e4ecf9dd6dc2f2e263c60e8de8e17aafedddb9.

Why a 1-bit model is the right vehicle for this

The Bonsai packs are not a niche. The 1-bit GGUF carries 636,606 downloads and 813 likes, the MLX 1-bit carries 1,954,689 downloads, and the ternary twin carries 557,054 downloads and 1,246 likes. All three were created on 4 July. Whatever the technical debates, the community downloaded this thing like a hit single.

Prism ML's published case is aggressive: true 1.125 bits per weight (one sign bit plus an FP16 scale amortized over a group of 128), ~3.9 GB deployed versus ~54 GB FP16, a 262,144-token context, and claimed speeds of 44 tok/s on an Apple M5 Pro and 104.8 tok/s on an H100, with math benchmarks that stay near FP16 while conventional 2-bit builds collapse. Those are upstream publisher claims from the Bonsai card, measured on their own infrastructure, and the discussion threads show not everyone gets the advertised speed. The important part for this post is the mechanism: the refusal-direction edit happened at full precision, and the 1-bit pack came after. That order is what makes this release worth a look.

The edit happened in FP16. The squeeze came after.

s3nh's card is unusually precise about the abliteration parameters: direction index 29.58, attention output projection max weight 1.43 at position 38.82, MLP down-projection max weight 1.39 at position 39.20. It used Heretic v1.4.0, the same tool family that has produced several models covered on this site. Measured on the FP16 edit: 6/100 refusals versus the original's 81/100, KL divergence 0.0033.

Then everything was thrown through the 1-bit grinder. A weight that survives abliteration as a precise BF16 value becomes, at Q1_0, either +scale or −scale. Refusal behavior is mediated by directions in activation space, and removing those directions is a projection on the weights; a projection followed by sign quantization is not the same projection. Nobody has published a refusal re-test on the Q1_0 pack. The Velum card itself lists every benchmark as TBD: general reasoning, coding, math, instruction following, MTP acceptance rate, all TBD. The 6/100 belongs to the FP16 edit, not to this artifact.

What is proven here?

The lineage is fully sourced: every hop in the chain above links to its repository, and the decensor's parameters and KL are published by the editor. What is not proven is the behavior of the final Q1_0 artifact: no refusal re-measurement, no capability rerun, no speed claim on the Velum files themselves. The decensor's 6/100 is a publisher measurement on the FP16 intermediate, honestly labeled as such.

How to run it

This is a llama.cpp model, not a vLLM job. The Q1_0 format needs the PrismML llama.cpp fork with the low-bit kernels (CUDA, Metal, CPU). No Ollama page exists for any model in this chain as of writing, so the direct route:

git clone https://github.com/PrismML-Eng/llama.cpp
cmake -B build -DGGML_CUDA=ON && cmake --build build -j
huggingface-cli download guell00/Velum-Unbound-Uncensored --include '*.gguf' --local-dir ./velum
./build/bin/llama-cli -m ./velum/VELUM-UNBOUND-Q1_0.gguf -p "Hello" -n 256 --temp 0.7 --top-p 0.95 -ngl 99

The drafter file is the DSpark speculative-decoding layer, a six-layer block-parallel transformer whose weights the Bonsai card says add roughly 0.5 GB at serving precision and give a measured 1.37x end-to-end decode speedup on the CUDA path. It is a lossless accelerator: verification preserves the target distribution, so the drafter affects speed, not output. Whether the Velum pack's drafter was trained against this exact uncensored target is not documented on the card; measure acceptance before trusting the speedup.

On cost: the primary estimate for a 27B dense class model is $5.45/hour managed (1 × H200). But a 3.9 GB weight file changes the math: this thing serves comfortably on a single L40S or even a consumer GPU, which lands closer to the $2.34/hour class, and on a laptop it costs only electricity. The H200 estimate is the conservative managed-serving figure.

The creator: guell00 (and the chain behind the edit)

guell00 is Miguel p r, a non-Pro Hugging Face user with six models and two followers. The profile shows a pattern: VELUM-Coder (4,990 downloads), Nexora-Qwen-Coder-4B (4,061), Nexora-Gemma-4-Coder (2,705). This is a small independent operator shipping local-first model packs, and the Velum card is explicit about the project's identity: "Local · Unbound · MTP", "🇧🇷 MADE IN BRAZIL", multilingual with Portuguese and English. No socials, no Ko-fi, no donation address on the card; the funding model is not documented and I will not invent one.

The people behind the actual edit deserve the credit. s3nh is a prolific Hugging Face editor, 273 models and 267 followers, and the decensor card reads like someone who knows the tooling cold. tommytracx operates the THOX.ai label ("Your AI. Your Data. Your Rules."), which published the intermediate Q1_0 quant and whose docs at docs.thox.ai describe an on-device stack. And the original Bonsai comes from Prism ML, whose whitepaper, demo repo and Discord sit behind the whole family. Four independent parties, one chain, zero coordination visible.

One honest line: safety filtering is significantly reduced across this lineage, which is the entire point of the chain, and the refusal measurement that exists was made on the FP16 intermediate, not on the 1-bit file you would actually download.

The idea, in plain words

What "1.125 bits per weight" means

Every number in a normal model is a precise value, 16 bits or more. A 1-bit model throws almost all of that away: each weight becomes just a sign, + or −, and a single shared scale value is used for every group of 128 weights to keep the numbers in a useful range. That is why a 27-billion-parameter model can shrink from ~54 GB to ~3.9 GB. Abliteration usually needs precise activation measurements to find and remove the refusal direction; this model had that done at full precision first, then got crushed down to signs. Whether the refusal removal still works after the crush is the honest open question.

Primary sources