license: other
tags:

• merge
• lora
• abliterated
• uncensored
• reasoning
• coding
• token-efficient
• mtp
• qwen3_8
• ai-red-team
• conversational
  base_model:
• ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP
• Shockem/Qwen3.8-27b-Terse-Coder-LoRA

Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder

Merged weights of ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP
with the Shockem/Qwen3.8-27b-Terse-Coder-LoRA
adapter (round 8, rank-16 DPO) baked in. The base is an abliterated Swift-Qwen3.8-27B — UkisAI's
reasoning-efficient fine-tune with orcarouter's single-direction refusal ablation applied to
Swift's own weights, vision tower untouched, MTP head kept and edited consistently [1]. The
adapter adds a conciseness pass targeted at coding traces. Single checkpoint, no LoRA plumbing.

⚠️ Read this before use. This model inherits the base's abliteration: its safety alignment
has been substantially removed, and it will comply with harmful, unethical, offensive, or
illegal requests that the original Qwen3.8-27B would refuse. It has no meaningful built-in
guardrails. It is intended for legitimate research — interpretability, AI-safety and
refusal-mechanism study, red-teaming, robustness evaluation, and controlled experiments. You
are responsible for how you use it, for complying with applicable law and the license, and for
everything it generates [1]. Do not deploy it to end users or in production without adding
your own safety, moderation, and abuse-prevention layers. Its outputs do not reflect the views
of the uploader, UkisAI, OrcaRouter, Shockem, or Qwen / Alibaba.

How it was made

• Merged in fp32: W + B @ A * (lora_alpha / r), with alpha 32 and r 16 (scale 2.0).
• Stored as bf16 using stochastic (unbiased) rounding with a fixed seed (0), so the merge is reproducible.
• MTP head and vision tower untouched by the adapter. The base's MTP head is abliterated
  consistently with the main model and the full vision tower is preserved, so self-speculative
  decoding and image understanding both keep working [1].
• All non-weight files (config, tokenizer, processor, index) copied from the base; the chat
  template is Shockem/froggeric-terse-coder, the one the adapter was evaluated with. Serving
  without it changes agentic behavior.

Why stochastic rounding

The adapter's weight deltas are deliberately tiny (‖Δ‖/‖W‖ ≈ 4e-4–1e-3), below bf16's
per-element resolution. The adapter card measures delta survival of only 31–61% under plain
bf16 rounding vs 94–99.9% in fp16. Stochastic rounding is unbiased — each element is rounded
up or down with probability weighted so its expected value equals the true merged value — so
the delta is preserved on average while keeping the checkpoint at the base's bf16 dtype and size.

Expected behavior and caveats

• This is the most layered artifact in the series — Qwen3.8-27B → Swift 1.0 → abliteration →
  Terse-Coder LoRA. Both of the two non-LoRA edits are themselves unmeasured for their
  interaction: the base card lists "whether Swift's shorter reasoning traces survive" abliteration
  as not evaluated [1], and the adapter's author has no results on any abliterated base. The
  conciseness effect should compound, but the magnitude here is unknown, and no independent
  benchmarks have been run on this artifact.
• Uncensored behavior is preserved in expectation. Abliteration is a weight edit, not
  data-level unlearning; a DPO on coding preference pairs should not restore guardrails. Treat
  this as expected, not verified.
• The adapter's author recommends runtime LoRA as the full-strength deployment form and
  measured a capability tax after merging on the stock Qwen base (70% → 60–62% on their
  held-out-40 after fp32-merge → fp16 → NVFP4 re-quant). This merge stores bf16 with no
  re-quant, so the tax should be smaller, but it is not zero.
• Do not load the Terse-Coder LoRA on top of this model. Double application over-shortens
  reasoning (63% pass with no_code failures in the adapter's testing).
• The adapter targets coding tasks with thinking enabled. It is a behavioral edit, not a
  knowledge edit — if a task needs long derivation, raise reasoning_effort as usual.
• Sampling, as for Swift and Qwen: temperature 1.0, top_p 0.95, top_k 20, min_p 0 [1].

How to use

Transformers

from transformers import AutoModelForImageTextToText, AutoProcessor
import torch

model_id = "vwdubb/Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

vLLM

vllm serve vwdubb/Swift-Qwen3.8-27B-Uncensored-MTP-Terse-Coder \
  --dtype bfloat16 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --port 8000

MTP self-speculative decoding (optional, head is included and untouched) [1]:

--speculative-config '{"method":"mtp","num_speculative_tokens":3}'

License

This model is a derivative of Swift-Qwen3.8-27B-Uncensored-MTP (Swift Open License v1.0) and
the Terse-Coder LoRA (Apache 2.0). The merged weights are released under the
Swift Open License v1.0: free for personal, research, educational, evaluation, and
commercial use by individuals and organizations with gross annual revenue up to
US$1,000,000; above that threshold, commercial use requires a separate Swift Enterprise
License from UkisAI [1]. Nothing in the Swift Open License limits your rights in
Qwen3.8-27B itself under Apache 2.0.

Acknowledgements

• ajgazin — the abliterated Swift base.
• UkisAI — Swift-Qwen3.8-27B.
• OrcaRouter — the refusal direction.
• Shockem — the Terse-Coder LoRA and chat template.
• Qwen Team (Alibaba Cloud) — Qwen3.8-27B.
• Arditi et al. — Refusal in Language Models Is Mediated by a Single Direction (2024) [1].