---Tags
colibri,qwen3.8,qwen3.8-27b,dense,abliterated,gated-deltanet,vision
Library_name
colibri
Base_model
Qwen/Qwen3.8-27B
Qwen3.8-27B (abliterated) for colibrì
An f16 container of an abliterated Qwen3.8-27B, packed in the format that the colibrì qwen36 engine reads.
This is not the official checkpoint, and it is not the original model. The weights are Qwen's, modified by abliteration (refusal-direction removal) by
What it is
Qwen3.8-27B is a dense model, not a mixture of experts: 64 layers, hidden size 5120, MLP width 17408, vocabulary 248320. Three of every four layers are Gated DeltaNet (linear attention) and every fourth layer is full attention. The checkpoint also carries a 27-layer vision tower, so the model accepts images.
colibrì treats the MLP of a dense checkpoint as an ungated shared expert (num_experts: 0), so there is nothing to stream from disk: all weights are read from RAM for every token.
What the conversion did
Done with tools/convert_qwen36.py from colibrì, with no quantization at conversion time.
Part Stored as
Trunk (attention, DeltaNet, MLP) f16
Embeddings, lm_head, final norm f16, in model-globals.safetensors
Vision tower f16, in model-vision.safetensors
MTP head not included (the converter skips mtp.* on purpose)
Size on disk: about 51 GiB in 66 safetensors files. Because the MTP head is skipped, the engine has no speculative decoding for this model.
The ebits: 4 field in qwen36_meta.json is written by the converter by default and has no effect here: it applies to experts, and this model has none.
Use
You need colibrì with the qwen36 engine and support for the dense Qwen3.8-27B (colibrì issue #1757).
bash
git clone https://github.com/JustVugg/colibri && cd colibri/c
make qwen36
./coli chat --model /path/to/this/repo --gpu none
The dense trunk is quantized by the engine when it loads, controlled by environment variables:
Setting Resident memory Notes
default (COLI_DENSE_BITS unset, int8) about 29 GB
COLI_DENSE_BITS=4 about 18-19 GB (estimate) int4 in blocks of 64, one f32 scale per block
COLI_DENSE_BITS=16 more than int8 no quantization, the container's own f16 values
With COLI_DENSE_BITS=4, COLI_DENSE_INT4 selects which components take int4 (lmhead, dnproj, dnout, attn, shexp, router). The engine source notes that the attention projections and the DeltaNet input projections cost the most perplexity, so a conservative setting is COLI_DENSE_INT4=lmhead,dnproj,dnout,shexp.
The expert-cache option does not apply to this model: leave --cap at 0.
Measured speed
Measured by the author on a laptop, CPU only (--gpu none), one conversation:
Setting Generation
int8 (default) 1.6 tok/s
COLI_DENSE_BITS=4 with COLI_DENSE_INT4=lmhead,dnproj,dnout,shexp 2.9 tok/s
A dense model reads all of its weights for every token, so decoding speed is set by memory bandwidth. Mixture-of-experts models read only the active experts and are much faster at the same total size.
Quality
Not benchmarked. No perplexity or task comparison against the original Qwen3.8-27B or against the unmodified abliterated checkpoint has been run, for either int8 or int4. Treat both as unverified.
Intended use and limits
This is an experimental model, not suited to consumer use. Abliteration removes the model's tendency to refuse; it does not remove errors, and it does not make the model safe to rely on. It is published for research on local inference, red-team and security research in controlled environments you own, creative writing, and synthetic data. You are responsible for what you do with its output.
License
<LICENSE: copy from the source checkpoint and from Qwen/Qwen3.8-27B, fill in>. The weights are Qwen's and the abliteration is by the author of the source repository; model licenses remain their own. The colibrì engine is Apache 2.0.
Source and credits
Original model: Qwen/Qwen3.8-27B, by the Qwen team.
Abliteration:
Engine and converter: JustVugg/colibri.