base_model: grimlee/Swift-1.5-Qwen3.8-27B-Abliterated
base_model_relation: quantized
license: other
license_name: swift-open-license-1.0
license_link: LICENSE
library_name: mlx
pipeline_tag: text-generation
tags:

• mlx
• kapitsa
• apple-silicon
• quantized
• oq6
• mtp
• dflash2
• speculative-decoding
• abliterated
• uncensored

Swift 1.5 Qwen3.8-27B Abliterated — oQ6, bf16 scales, MTP + DFlash2 (for Kapitsa)

An abliterated Swift 1.5 Qwen3.8-27B, quantized to oQ6 and packaged so that
Kapitsa can serve it at full speed: bf16 quantization metadata, the model's
own MTP head in a sidecar, and a DFlash2 drafter bundled.

What abliterated means here

The refusal behaviour has been removed from these weights. This model will
answer requests that the original Swift 1.5 would decline. That is the
purpose of an abliterated model, and it is worth being deliberate about where
you run one and who can reach it — by default kapitsa serve listens on your
network, so --private or --api-key is a sensible habit for this model in
particular.

The abliteration was not performed here. It comes from
grimlee/Swift-1.5-Qwen3.8-27B-Abliterated,
whose ABLITERATION_METADATA.json ships unmodified in this repo: single-
direction refusal abliteration, lambda 1.2, 131 tensors rewritten. Its own
evaluation reports refusals on XSTest-30 falling from 25/30 to 0/30, with
MMLU-style accuracy unchanged (40/40 both ways, zero prediction changes). None
of that was re-verified here.

What this repo adds is the quantization and the Kapitsa packaging.

For the unmodified model, use
jongstra/Swift-1.5-Qwen3.8-27b-oQ6-bf16-mtp-dflash2-kapitsa.
It is built with an identical recipe — the same 455/50 allocation — so the two
differ only by the abliteration.

Requirements

• Kapitsa 0.2.25 or newer. Earlier versions reject a 6-bit MTP sidecar with
  mtp.layers.0.self_attn.q_proj: unsupported quantization.
• Apple Silicon with about 28 GB free for the weights, the drafter and one
  conversation slot. Measured on an M2 Ultra at 22.5 GB of weights plus 1.86 GB
  of resident drafter, plus KV.

How to download and run it

Kapitsa's catalog is built into the binary, so this model cannot be selected
by name — not with kapitsa serve --model <name>, not from the interactive
picker, and not through POST /v1/models/switch. It is loaded by folder
path.

The short way

hf download jongstra/Swift-1.5-Qwen3.8-27b-Abliterated-oQ6-bf16-mtp-dflash2-kapitsa \
    --local-dir ~/models/swift15-abl-oq6-kapitsa

kapitsa serve --model ~/models/swift15-abl-oq6-kapitsa --model-id swift15-abl --port 8000 --private

So that kapitsa list shows it

This repo ships a .kapitsa/manifest.json, which is what kapitsa list reads:

hf download jongstra/Swift-1.5-Qwen3.8-27b-Abliterated-oQ6-bf16-mtp-dflash2-kapitsa \
    --local-dir ~/.cache/kapitsa/models/swift-1.5-qwen3.8-27b-abliterated-oq6

kapitsa list
#   swift-1.5-qwen3.8-27b-abliterated:oq6   28.16 GB  weights + MTP head + DFlash2 drafter

It will be listed, but still has to be started by path — the name is not
resolvable.

A successful start logs:

MTP head loaded in … (dense, trunk hidden pre-norm, drafts over the full vocabulary)
loaded …: 64 layers, dense, 22.5 GB mapped; weights 6-bit group 64 87%, 8-bit group 64 13%
DFlash2 drafter …/dflash2: 5 layers, hidden 5120, 4-bit projections, 1.86 GB resident
engine ready: context 262144, KFlash speculative decoding (DFlash2 lattice, tree verify), images on

If the engine reports MTP speculative decoding instead, the dflash2/ folder
did not download — it still works, about 30% slower.

What is in here

• oQ6: oMLX's sensitivity-driven mixed precision, not uniform 6-bit. 505
  quantized linear modules — 455 at 6-bit group 64, 50 at 8-bit group 64 —
  allocated from the per-layer scores in oq_sensitivity_map.json. Built with
  oMLX 0.7.3 at dtype="bfloat16", preserve_mtp=True.
• The model's own MTP head, 6-bit group 64, copied verbatim into
  mtp.safetensors as mtp.*.
• A DFlash2 drafter in dflash2/, from
  incoai/Qwen3.8-27B-DFlash2
  (Apache-2.0), unmodified.
• Vision tower preserved — the model reads images.

Detail in CHANGES.md, conversion_metadata.json
and ABLITERATION_METADATA.json.

Measured on an M2 Ultra (60-core GPU, 64 GB)

Kapitsa 0.2.25, one slot, 700-token generations, median of three runs:

All three are within run-to-run variance of each other. Single-request
micro-benchmarks on one machine, not a rigorous evaluation.

Licence and attribution

Derivative Work of ukisai/Swift-1.5-Qwen3.8-27b under the Swift Open
License v1.0 — see LICENSE, NOTICE and the source card
copied as SOURCE_MODEL_CARD.md. Note the licence's
commercial use limitation and its revenue threshold. Modifications are listed
in CHANGES.md as §4(b) requires, separated into those made here
and those inherited.

Abliteration by grimlee. Drafter by
incoai, Apache-2.0. Quantized with
oMLX 0.7.3.