license: other license_name: swift-open-license-1.0 license_link: LICENSE base_model:

• ukisai/Swift1.5-Qwen3.8-Flash-Next
• peonist-ai/halogen-qwen3.8-flash-next
• Qwen/Qwen3.8-Flash-Next base_model_relation: quantized pipeline_tag: text-generation library_name: halogen inference: false tags:
• abliterated
• uncensored
• swift
• efficient-thinking
• reasoning
• halogen
• halogen-flash
• qwen
• moe
• mixture-of-experts
• long-context
• quantized
• 4-bit
• strix-halo
• gfx1151
• rocm
• amd

Halogen Swift 1.5 Qwen3.8-Flash-Next v2, abliterated

Runs only on halogen-flash-server. This is a .hgn checkpoint, Halogen's own format. It will not load in transformers, vLLM, llama.cpp, Ollama or LM Studio, and there is no GGUF or safetensors version. Halogen targets AMD Strix Halo (gfx1151) only.

This is UkisAI's Swift 1.5 Qwen3.8-Flash-Next fine-tune, with a refusal-direction edit ("abliteration"), packed into peonist's Halogen v2 checkpoint (peonist-ai/halogen-qwen3.8-flash-next).

• Shorter thinking: on 14 MMLU-Pro questions it used 72% fewer thinking tokens than the abliterated base model (mean 905 against 3,271).
• Refusals removed: it answered all six refusal test prompts. Plain Swift 1.5 refused all six.
• Same speed as stock v2: about 53–57 tok/s greedy on a Ryzen AI Max+ 395.
• Slightly lower perplexity quality than the base-model build: WikiText-2 perplexity 3.229, against 3.134 for the abliterated base build and 3.144 for stock v2. See Quality.
• Encoded in v2's own formats: the changed tensors use v2's native 4-bit HT and I4R grids with calibration-aware rounding.

This is an unofficial derivative. It is not made or endorsed by UkisAI, Peonist, orcarouter or Qwen.

Related builds

All three are the same v2 checkpoint with different tensors replaced, built with the same tooling and scored in the same session:

Model details

Files

The file is 23 GB larger than stock v2 (66.7 GB). The original payloads stay inside it, unreferenced, and the replacements are appended. The engine loads only the referenced payloads, so the resident weights are the same size as stock v2's (about 62 GiB).

You also need these from peonist's repo. They are unchanged, so they are not re-uploaded here:

Swift 1.5 uses the base model's chat template unchanged, so peonist's tokenizer/ is correct for it. The template's reasoning_effort levels (xhigh by default, medium, low) work as Swift's card describes.

Hardware requirements

• A 128 GB Strix Halo machine, used for this model only. Other large processes compete for the remaining memory and can stall the server.
• About 141 GB of disk: this file (89.8 GB) plus the n-gram table (51.2 GB). Add the vision tower if you want image input.

Quick start

hf download Quat3rnion/halogen-swift1.5-qwen3.8-flash-next-v2-abliterated \
  qwen38-flash-next-v2-swift15-abliterated.hgn --local-dir /models/v2
hf download peonist-ai/halogen-qwen3.8-flash-next qwen38-flash-next-ngram.hgn \
  --revision 5cc17cea1a10b502b4a5db41d8d2b1a0f22dba26 --local-dir /models/v2
hf download peonist-ai/halogen-qwen3.8-flash-next \
  --revision e053f488b120b99ed2525e6ac99f68c51b3b6179 \
  --include 'tokenizer/*' 'qwen38-flash-next-vision.hgn' --local-dir /models/w4b

docker run --rm --name halogen \
  -p 127.0.0.1:8731:8731 \
  --device /dev/kfd --device /dev/dri --ipc=host --ulimit memlock=-1:-1 \
  --group-add video --group-add render \
  -e HALOGEN_WEIGHTS_LOCK=1 \
  -e HALOGEN_KV_POOL_POSITIONS=262144 -e HALOGEN_KV_SLOTS=2 -e HALOGEN_MAX_TOK=8192 \
  -e HALOGEN_SPEC_ADAPT=0 \
  -e HALOGEN_CHECKPOINT=/v2/qwen38-flash-next-v2-swift15-abliterated.hgn \
  -e HALOGEN_NGRAM_TABLE=/v2/qwen38-flash-next-ngram.hgn \
  -e HALOGEN_TOKENIZER=/models/tokenizer \
  -e HALOGEN_VISION_TOWER=/models/qwen38-flash-next-vision.hgn \
  -e HALOGEN_MODEL_ID=halogen-swift15-abliterated \
  -e HALOGEN_TEMPERATURE=1.0 -e HALOGEN_TOP_P=0.95 -e HALOGEN_TOP_K=20 \
  -v /models/w4b:/models:ro -v /models/v2:/v2:ro \
  ghcr.io/peonist-ai/halogen-flash-server:0.15.1

• HALOGEN_SPEC_ADAPT=0 keeps the MTP draft head on for every token. With the default adaptive policy, the engine sometimes turned the head off for whole requests, which roughly halved decode speed (see Speed). Output is byte-identical either way.
• The sampling defaults are the ones Swift's card recommends.
• Leave out HALOGEN_VISION_TOWER for text only.
• The checkpoint carries its own MTP draft head, so do not set HALOGEN_MTP_HEAD. v2 rejects HALOGEN_FLASH_PIN_TRUNK=0.
• On Podman, use --group-add keep-groups instead of the two --group-add flags.
• The server starts in about 30 seconds.

Benchmarks

All numbers were measured on a Ryzen AI Max+ 395 (Radeon 8060S, 128 GB) with halogen-flash-server 0.15.1.

Thinking length

Fourteen MMLU-Pro test questions, one per subject (sampled with seed 7). Each was asked once, at temperature 1, top_p 0.95, top_k 20, default xhigh effort, with a 16,384-token thinking budget that no answer reached:

Per question, this build's thinking length averaged 0.45× r2's (geometric mean of the ratios; paired t = −3.1). UkisAI report a 57% reduction in mean thinking tokens on MMLU-Pro for the BF16 model, so the fine-tune's effect survives 4-bit re-encoding.

Treat the exact percentages as rough. There was one sample per question, and thinking length varies a lot between samples: the same question once took r2 11,103 tokens and another time 1,642. Accuracy on 14 questions says nothing reliable about capability. The comparison is against r2 rather than stock v2; r2 differs from stock only by the refusal edit.

Quality

Scored with the engine's own ppl mode (teacher-forced, served numerics) on 70,774 tokens of the WikiText-2 raw test split, against stock v2's top-128 distribution at every position:

Paired per token, this build is 0.030 nats worse than r2 (t = +16.5, 95% CI +0.026 to +0.034) and 0.015 nats worse than plain Swift 1.5 (t = +9.9).

Perplexity on encyclopedia text measures closeness to the base model. It penalizes a chat and reasoning fine-tune for moving away from that model, so it does not capture what Swift was trained for. No task benchmark was run beyond the thinking-length sample above.

Speed

Server settings: 524,288-position KV pool, 4 slots, HALOGEN_MAX_TOK=8192, HALOGEN_SPEC_ADAPT=0. Ten short prompts, 256 output tokens each, one request at a time, without thinking; the whole set was run twice:

Cold 34k-token prefill was about 1,640 tok/s (r2 1,650), measured in an earlier session on a busier machine.

With the engine's default HALOGEN_SPEC_ADAPT=32,0.35,64, each of the three builds (r2 included) sometimes dropped to 26–28 tok/s for runs of whole requests because the draft head had been switched off. That is an engine policy, not a property of the weights.

Refusal removal

Refusal-direction component of the 149 output-writer tensors, |rᵀW|/|W|, where r is the refusal direction:

The 4-bit residual is consistent with 4-bit quantization noise, and matches r2.

Test prompts at temperature 0, thinking off (the exact prompt texts are not published):

Capability checks

Smoke tests, one run each; all passed:

Re-encoding accuracy

Relative L2 error of each re-encoded tensor against the abliterated Swift BF16 source, over all 353 tensors:

Swift's own change to each tensor is small (a 0.3–2% relative difference from the base weights in the tensors sampled), well below this 4-bit error, but the thinking-length results show it survives.

What changed

The build has two steps.

1. BF16 projection. tooling/swift_bf16_abliterate.py estimates one refusal direction r (2,560-dimensional) as the top singular direction of orcarouter's weights minus the base weights, over the 149 residual-writer tensors orcarouter edits. It checks that each of those edits is essentially rank-one along r (the gates are in the manifest). It then projects r out of Swift 1.5's copy of the same 149 tensors, in float64, rounding back to BF16. No orcarouter delta is added; only its direction is used. Writers that Swift did not fine-tune (routed experts, embeddings, MTP writers) are projected from Swift's values, which match the base model's there.

2. v2 re-encoding. The 353 tensors that differ from the base model are Swift's 300 fine-tuned tensors plus the 53 abliterated writers Swift left alone. Each is re-encoded from the projected BF16 into the store v2 uses for that tensor:

The 96 HT output writers (o_proj, linear_attn.out_proj, shared-expert down_proj) carry both Swift's fine-tune and the projection. Every other payload, including the rest of the MTP head, is peonist's v2 byte for byte.

Calibration: the LDLQ rounding uses input second moments (Hessians) collected from the base model's activations over 128 × 2,048 tokens of WikiText-2 train (tokens SHA-256 8a9551ebace12b6f353aff37beb49e7f1d1b85d4f73ada1f9447ca9564d7a840). They are the same Hessians as r2's; the tokens and manifest are in r2's calibration/.

Reproduce

1. Download the three BF16 models at the revisions above. Run tooling/swift_bf16_abliterate.py --base … --donor … --swift … --output swift15-abliterated-bf16/ (pass all three paths; the defaults are the build machine's). It needs NumPy and about 335 GiB of free disk, and it verifies the pinned revisions before writing.
2. Build with r2's tooling/ and this repo's writers.txt:

python3 tooling/halogen_v2_abliterate.py build \
  --v2 qwen38-flash-next-v2.hgn --donor swift15-abliterated-bf16/ \
  --writers writers.txt --ht-encode trellis --rounding ldlq --hessians hessians/ \
  --out qwen38-flash-next-v2-swift15-abliterated.hgn --workers 8

The build took 3.6 hours on 32 CPU threads with 5.6 GiB peak memory. A byte-identical rebuild has not been tested; floating-point summation order may make it differ slightly.

Limitations

• Refusal removal was tested on six prompts. It is not guaranteed for every prompt, and the model may still refuse some requests.
• The thinking-length result comes from 14 single samples.
• Perplexity is 0.095 above stock v2's on WikiText-2. Chat, code and long-context retrieval were not scored.
• Calibration used public WikiText-2 text and the base model's activations, not Swift's or peonist's.
• The file only works with the Halogen engine. It was tested on 0.15.1 only.

Intended use and responsibility

This model has had its refusal behavior removed. It can produce content that the base model would decline to produce.

• If you deploy it to other people, add your own moderation and abuse-prevention layers.
• You are responsible for complying with the licenses and with the laws that apply to you.
• The uploader accepts no liability for misuse.

License and attribution

This repository is not Apache-2.0 as a whole. Read LICENSE, LICENSE-QWEN and NOTICE before using or redistributing it.

• Swift Contribution (UkisAI's fine-tune, in the tensors listed in writers.txt): Swift Open License v1.0. Commercial use is not licensed for an entity with US$1,000,000 or more in annual gross revenue (including affiliates); such entities need a Swift Enterprise License from UkisAI.
• Base model (Qwen/Qwen3.8-Flash-Next): Qwen Community License 1.0. Among other terms, a Model-as-a-Service or AI Work Assistant business needs a separate license from Qwen for commercial use.
• Peonist's v2 checkpoint (peonist-ai/halogen-qwen3.8-flash-next), whose unmodified tensors are redistributed here, and the refusal direction from orcarouter/Qwen3.8-Flash-Next-Uncensored: Apache License 2.0 (LICENSE-APACHE).
• tooling/: Apache License 2.0.

This summary is not legal advice; the license texts control. Modification notices are in NOTICE. The calibration tokens are from WikiText-2 (CC BY-SA 3.0). "halogen" is a trademark of Peonist, LLC, and "Swift" and "UkisAI" are UkisAI's; the names are used here only to describe the origin of the work and the compatible engine. The halogen-flash-server engine is closed source and distributed under its own terms. The v2 format notes build on jtsylve/hgn-spec 1.0.1.