Swift 1.5 Flash-Next — Heretic V2 (HGN v2)

Qwen 3.8 Swift 1.5 Flash-Next, Heretic V2 (uncensored, tuned for coding and reverse engineering), converted from trubisky's IQ4_XS-plus GGUF into the HGN v2 weight format used by halogen-flash-server (AMD Strix Halo, gfx1151).

Uncensored model. Refusal behavior has been deliberately removed. You are solely responsible for how you use it. Comply with all applicable laws and the licenses below.

Naming: "Heretic V2" is the version of the abliterated weights. "HGN v2" is the format profile of this conversion. They are unrelated version numbers.

These weights are in .hgn format and will not load in transformers, vLLM or llama.cpp. They are meant to be mounted into the halogen-flash-server container.

Status (read this first)

This is a community conversion. It has not been qualified by the authors of any upstream project.

Check	Result
Conversion (`bf16-to-hgn convert`)	Completed and published: 1,658 source tensors to 2,349 output tensors, manifest included
`bf16-to-hgn verify` (source hashes, layout, decoded values)	Completed: source hashes and layout checked, per-tensor decoded error recorded (see below)
Loads and generates in halogen-flash-server	Loads and serves in 0.18.22 (startup 12.5 s); quality benchmarks pending
CUDA output compared byte-for-byte with the CPU backend	Not done. The converter documents CUDA as qualified on sm_89; this run used sm_86 (RTX 3090)

If you test it, please report results on the discussion tab.

Conversion error (from `bf16-to-hgn verify`)

Relative error energy of the output's decoded values against the decoded GGUF source:

Tensor family	Error
Routed experts (I4R)	0.94% in every layer
Dense projections, LM head, embeddings (HT4)	0.6–1.1%
Hyper-connection mixers (Q6G64)	~0.06%
N-gram table (FP8)	0.07%
Norms, router, SSM state, vision tower	exact

This measures the second rounding step only, not distance from the original BF16 weights.

Speed (halogen-flash-server 0.18.22, Strix Halo, 124 GiB)

Short prompts, 512 output tokens, MTP on: ~56 tok/s single user; 93 tok/s aggregate at 4 concurrent; ~1,800 tok/s prefill at 6.1K tokens. No quality benchmark has been run yet.

Files

File	Size	Purpose
`model.hgn`	66,687,683,040 B (62.1 GiB)	Checkpoint, identity `qwen3.8-flash-next-v2`. Includes the MTP draft head
`qwen38-flash-next-ngram.hgn`	51,200,246,084 B (47.7 GiB)	The 51B-parameter n-gram lookup table (FP8). Required; keep beside `model.hgn`
`qwen38-flash-next-vision.hgn`	897,916,416 B (0.84 GiB)	Vision tower (BF16). Optional; without it the engine is text-only
`config.json`		Model configuration
`tokenizer/`		Flat tokenizer directory (also contains `config.json`)
`conversion-manifest.json`		Converter manifest: policy, spec version and per-file details

Total about 111 GiB. The model needs about 62 GiB resident plus the memory-mapped n-gram table.

How it was made

Qwen/Qwen3.8-Flash-Next                      base architecture
  -> ukisai/Swift1.5-Qwen3.8-Flash-Next       Swift 1.5 post-train
  -> jeffpeng3/Swift-1.5-Qwen3.8-Flash-Next-heretic   Heretic V2 abliterated weights
  -> trubisky/Swift-1.5-Flash-Next-Heretic-ReverseEngineer   IQ4_XS-plus GGUF (94.7 GB)
  -> this repo                                HGN v2 via bf16-to-hgn

Converter: Quat3rnion/bf16-to-hgn, profile v2, policy bf16-hgn-v2-rtn-nf4-signed-rms-trellis-v1, --ht-beam 128, seed 0. Format per hgn-spec 1.1.0.
Backend: CUDA (--ht-backend cuda --i4r-backend cuda --q4cp-backend cuda) on an RTX 3090. The conversion passes took 35 minutes 35 seconds.
MTP head: a GGUF carries no MTP head the engine can run, so the 31 mtp.* tensors were copied byte-for-byte from the published qwen38-flash-next-v2.hgn of peonist-ai/halogen-qwen3.8-flash-next into a standalone sidecar and passed to the converter with --mtp. The head is the stock one, trained against the stock trunk and not against the Heretic weights.
Vision tower: the published qwen38-flash-next-vision.hgn from the same repo (333 BF16 tensors), passed with --vision.
Assets: config.json from the same author's HGN checkpoint; tokenizer files and chat template from the stock Flash-Next tokenizer directory.

Reproduce:

bf16-to-hgn convert --source-format gguf \
  --source Swift-1.5-Qwen3.8-Flash-Next-Heretic-V2-IQ4_XS-plus-00001-of-00006.gguf \
  --assets assets --mtp mtp.hgn --vision qwen38-flash-next-vision.hgn \
  --profile v2 --output swift-v2 --threads 8 --ht-beam 128 \
  --ht-backend cuda --i4r-backend cuda --q4cp-backend cuda

Precision: what this is and is not

This is not a quantization from BF16. The source is an already-quantized IQ4_XS-plus GGUF. The converter decodes its values and re-encodes them into the HGN v2 formats, which adds a second rounding step on top of the original quantization. The converter makes no claim of BF16-equivalence.

The HGN v2 profile assigns formats by tensor family (I4R routed experts, HT4 trellis dense matrices and LM head, Q6G64 hyper-connection mixers, Q8G64 for the MTP head's dense projections, BF16 for the router, norms and vision). The per-layer bit allocation of the GGUF (for example 8-bit down projections in the last layers) is not carried over as a format choice; only the values it produced are re-encoded.
Expect quality close to the GGUF but not identical to it. No perplexity or benchmark comparison against the GGUF has been run for this release.

Speculative decoding

The checkpoint includes the MTP head. Per the halogen documentation, speculative decoding is byte-identical to serial greedy decoding, so a poorly matched head costs speed, not correctness. Because this head was not trained against the Heretic weights, expect a lower draft acceptance rate than with the stock model. It can be disabled per request.

Use with halogen-flash-server

Requires an AMD Strix Halo machine (gfx1151) with enough unified memory for roughly 62 GiB of weights plus the memory-mapped table. Mount the downloaded directory read-only, as for any halogen checkpoint:

hf download johnlockejrr/Swift-1.5-Flash-Next-Heretic-V2-HGN --local-dir ~/halogen-models

podman run --rm -p 8731:8731 \
  --device /dev/kfd --device /dev/dri --group-add keep-groups \
  --ipc=host --ulimit memlock=-1:-1 \
  -e HALOGEN_CHECKPOINT=/models/model.hgn \
  -e HALOGEN_VISION_TOWER=1 \
  -v ~/halogen-models:/models:ro \
  ghcr.io/peonist-ai/halogen-flash-server:<version>

Notes:

Replace <version> with a current server release. Setting HALOGEN_VISION_TOWER=1 is optional and only needed for images.
HALOGEN_CHECKPOINT points at model.hgn because this checkpoint is not named qwen38-flash-next-v2.hgn. This invocation is untested; see the server documentation if the engine does not find the n-gram file beside the checkpoint.
On Docker, replace --group-add keep-groups with --group-add video --group-add render.
The server exposes an OpenAI-compatible API on port 8731.

Credits

Base architecture: Qwen/Qwen3.8-Flash-Next
Post-train: ukisai/Swift1.5-Qwen3.8-Flash-Next
Heretic V2 abliterated weights: jeffpeng3/Swift-1.5-Qwen3.8-Flash-Next-heretic
GGUF quantization recipe, imatrix and quants: trubisky/Swift-1.5-Flash-Next-Heretic-ReverseEngineer. The IQ4_XS-plus variant is Unsloth's UD-IQ4_XS recipe with KL-guided step-ups of ffn_gate_exps and ffn_up_exps on layers 41, 42, 44, 45, 46 and 47, quantized against a fused imatrix weighted toward coding, reasoning, agentic tool use, design, spatial reasoning and reverse engineering
Fusion baseline imatrix: the AtomicChat team
HGN format specification: jtsylve/hgn-spec
Converter: Quat3rnion/bf16-to-hgn
Engine, stock HGN checkpoint, MTP head and vision tower: peonist-ai (halogen-flash-server)
HGN conversion of the GGUF: johnlockejrr

This repository is an independent community conversion and is not affiliated with or endorsed by any of the above.

License

Derived from Qwen3.8-Flash-Next through the Swift 1.5 post-train and the Heretic V2 weights. It inherits the Swift Open License v1.0 and the Qwen Community License. The MTP head and vision tower come from peonist-ai's checkpoint, which is governed by the license of the original model. Nothing here grants rights that the upstream licenses do not. Verify the terms yourself before commercial use.

The halogen-flash engine is separate software distributed under its own terms. "halogen" is a trademark of Peonist, LLC; see TRADEMARKS.md in the server repository for permitted use.