Halogen 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.

Current native v2 release (2026-10-11)

qwen38-flash-next-v2-abliterated.hgn is now a complete native-v2 conversion of the original OrcaRouter BF16 source, revision 8336e613ea508b13c2159bd0f68965d97a606b95. It is not the r1/r2 writer transplant into peonist's checkpoint, and does not reuse stock HGN payloads or stock sidecars. The filename is retained so existing exact-file download links still resolve.

This native v2 release was built with Quat3rnion/bf16-to-hgn.

• Main file: 66,687,678,432 bytes; SHA-256 b5bd925fa2b44343be4a2cd328cae6b57ce7ffa2fe2cc56281717377e5a1ce8a.
• Converter commit: 04276213faff928aed70a7ac84c12e67460ea047; immutable binary SHA-256 52e83980ba24913d44788c92f3c484619303f576b394cee8ae0760f0a23efdf3.
• Conversion: original HF BF16, profile v2, HIP HT backend, seed 0, 2 threads, trellis beam 128; policy bf16-hgn-v2-rtn-nf4-signed-rms-trellis-v1. This release is not the historical LDLQ-calibrated transplant.
• Full HIP source verification and CPU-native format checks of the main, n-gram and vision outputs passed before publication. These are offline checks, not GPU inference, generation-quality or performance results.
• The native bundle derives its n-gram table, vision tower, embedded MTP head and nine copied assets from the same pinned BF16 source. The MTP head is embedded; no separate stock MTP head was used.

Companion provenance and availability

This replacement publishes only the main checkpoint and this provenance/ license correction, not the complete converter output directory. The source-derived n-gram and vision files and copied assets are not newly uploaded here.

The historical download recipe below pins peonist's n-gram to 5cc17cea1a10b502b4a5db41d8d2b1a0f22dba26 and tokenizer/vision to e053f488b120b99ed2525e6ac99f68c51b3b6179. Those dependencies are needed for the historical r2 recipe; their interchangeability with this conversion has not been established. In particular, the planned source-derived n-gram is 51,200,246,084 bytes, whereas that stock file is 51,200,246,144 bytes, so whole-file identity cannot be assumed. The source-derived vision file is planned at 897,916,416 bytes, the same size as stock, but equal size is not proof of equal content. Pinned repository metadata has matching blob identities for the six tokenizer assets (chat template, generation config, merges, tokenizer JSON/config and vocabulary); that is not a claim that all bundle companions are identical or that inference has been validated.

Use source-matched companions from a verified local conversion; do not treat the retained stock-companion quick start as validated for the new file. No new stock-companion compatibility, perplexity, refusal, speed, vision, tool-call, context or resident-memory result is claimed for this release.

License and rollback

The pinned original source's actual LICENSE contains Qwen Community License 1.0, despite its model card's Apache-2.0 metadata. The full notice is retained verbatim in LICENSE-QWEN; consult that notice for its conditions. This is a modified/re-encoded distribution of that source. The existing Apache LICENSE, tooling/LICENSE and ht43-tooling/LICENSE are retained; the tooling license is not changed.

Historical r2 remains downloadable at revision e7783c5822dfda2785a0f5b435abf54df89fe8ee with SHA-256 7bbac319851b500639462133085156b4f998dec562f511e65ca1c14d7af5163e and size 88,737,777,792 bytes. No history rewrite is used. The experimental HT43 checkpoint, calibration data and tooling are unchanged.

Historical r1/r2 transplant overview

The descriptions and benchmark results below are historical r1/r2 results, not results for the new native conversion, unless explicitly marked current or HT43. The retained tooling and calibration reproduce those historical builds, not the new native conversion.

This is peonist's Halogen v2 checkpoint for Qwen3.8-Flash-Next (peonist-ai/halogen-qwen3.8-flash-next), with the refusal-direction edit ("abliteration") from orcarouter/Qwen3.8-Flash-Next-Uncensored applied. It runs on halogen-flash-server 0.15.0 on AMD Strix Halo (gfx1151), at the same speed as stock v2.

• Near-stock quality: perplexity on WikiText-2 test is 3.134, against 3.144 for stock v2. Its next-token distribution is close to stock's (KL 0.060, same top token 92.2% of the time).
• Same speed as stock v2: about 52–54 tok/s greedy on a Ryzen AI Max+ 395.
• Refusals removed: stock v2 refused all three original test prompts; this checkpoint answered all of them, and three more.
• Only 149 tensors changed: everything else is byte-identical to peonist's v2. It contains no other fine-tune.
• Encoded in v2's own formats: the changed tensors use v2's native 4-bit HT grid and calibration-aware rounding.
• Open tooling: the scripts that built it are in tooling/, and the calibration input is in calibration/.

Versions

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

Related builds

The same v2 checkpoint is also available with UkisAI's Swift 1.5 reasoning-efficiency fine-tune, built with the same tooling and scored in the same session:

The Swift builds think 58–72% less on that sample at the same speed, and trade some perplexity for it. They are under the Swift Open License v1.0, which limits commercial use by larger companies; see their cards.

Experimental HT43 alternative

v2 remains the recommended default. qwen38-flash-next-ht43-abliterated.hgn is a separate, optional experimental checkpoint published in this repository, not a replacement for the existing v2 file. The historical v2 download and server commands below are retained for r2; see the current release's companion caveats. Downloading either filename does not select or replace the other checkpoint automatically; checkpoint selection is explicit through HALOGEN_CHECKPOINT.

This non-Swift build applies the same orcarouter abliteration donor used by the v2 r2 release, revision 8336e613ea508b13c2159bd0f68965d97a606b95, to the pinned HT43 checkpoint. It does not add Swift's fine-tune. The 48 routed down-projection writers were independently encoded in native HT4/HT3 with trellis beam 128 and calibrated LDLQ rounding (damping 0.01), using base-model WikiText-2-train Hessians. Compatible non-expert encoded writers were reused only after complete stock payload and transform-plane equality checks, then composed into a compact output.

What HT43 changes

The source is peonist's qwen38-flash-next-ht43.hgn at revision 61d3a32930e6e78d996f72f7256b5dd84ada2144. Its tensor table has the same 2,015 names and shapes as v2. The trunk routed experts use HT4 in even layers and HT3 in odd layers, rather than v2's 4-bit I4R format; 290 hyperconnection tensors change from q6g64 to q8g64. This changes quantization, not the architecture or native context length.

The published derivative is 57,618,253,824 bytes (53.66 GiB), versus upstream v2's 66,687,678,432 bytes (62.11 GiB): about 13.6% smaller. Its SHA-256 is 2beaad7b6d5107d7abfa866b33b47287160a68b9fc45b86020dbaa884955329c. The upstream HT43 input is 57,618,253,792 bytes, SHA-256 12e8e91b06dfd3bad4b68cdd275138733da64542a254612b116e8bd9fe77fdd8. These are checkpoint file sizes, not measured resident memory. The historical v2 r2 derivative is larger on disk because it retains unreferenced original payloads; its referenced weights remain about 62.11 GiB. HT43 changes the live quantization representation, rather than merely discarding obsolete bytes. Actual resident-memory savings and the quality tradeoff have not been measured. Do not apply v2 benchmark numbers to HT43.

Less weight memory can leave more room for the KV cache, including a larger shared pool or more concurrent native-context requests. It does not itself extend the model's native 262,144-token context per request. YaRN context extension is a separate opt-in and needs its own quality, long-context and memory validation; no extended-context result is claimed here. A larger KV pool alone is not proof of a longer usable context.

Optional download and selection

Download the published experimental file separately without changing the v2 default:

hf download Quat3rnion/halogen-qwen3.8-flash-next-v2-abliterated \
  qwen38-flash-next-ht43-abliterated.hgn --local-dir /models/ht43

Keep the n-gram table, tokenizer and optional vision downloads described below. To opt in deliberately, use an HT43-capable Halogen engine, add a read-only mount -v /models/ht43:/ht43:ro to the server command, and replace only the checkpoint argument with:

-e HALOGEN_CHECKPOINT=/ht43/qwen38-flash-next-ht43-abliterated.hgn

Do not assume that the older engine version in the v2 quick start supports HT43. Engine 0.15.3's metadata inspector recognizes HT3, but that inspection is not a GPU load or inference test of this derivative. Keep the other v2 settings and dependencies unchanged for an initial comparison; increase cache allocation only after measuring available memory and exercising the intended request load.

Tooling directories

• tooling/: the original normal v2 build and verification tooling, retained unchanged. The v2 reproduction commands below use it.
• ht43-tooling/: separate experimental HT43 tooling, provided alongside the normal v2 tools, not as their replacement.

Validation status and licensing

The completed 2026-10-05 CPU build, composition and numerical-verification receipts all report successful completion. The existing verify log reports 149 changed tensors, preserving each writer's native store and precision, and tabulates decoded-weight error and refusal-direction projection against the BF16 donor. This is offline weight-level numerical verification, not model inference or evidence of refusal behavior. The published file's complete SHA-256 was independently rechecked before upload on 2026-10-06.

This HT43 derivative has not been GPU benchmarked or inference validated. No HT43 perplexity, refusal, speed, vision, tool-call, long-context or resident-memory results are claimed. The model details, benchmarks and reproduction instructions elsewhere in this card describe v2, unless explicitly labeled HT43. Publication does not imply serving acceptance.

The existing license terms, attribution and responsibility guidance in this card also apply to this experimental variant. Choosing HT43 does not remove or relax any license restriction.

Model details

Historical r1/r2 section; not validation of the current native conversion.

Files

The historical r2 file is 22 GB larger than stock v2 (66.7 GB). The original writer 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.

You also need these from peonist's repo. These are historical r1/r2 dependencies, not newly uploaded here; see the current release's companion caveats:

Hardware requirements

Historical r1/r2 section; not validation of the current native conversion.

• A 128 GB Strix Halo machine, used for this model only.
  • At a 262,144-position KV pool with 2 slots, the server container used 64.8 GiB. The n-gram table is memory-mapped on top of that.
  • Peonist's advice applies here too: other large processes compete for the remaining memory and can stall the server.
• About 137 GB of disk: this file (88.7 GB) plus the n-gram table (51.2 GB). Add the vision tower if you want image input.

Quick start

Historical r1/r2 section; not validation of the current native conversion.

hf download Quat3rnion/halogen-qwen3.8-flash-next-v2-abliterated \
  qwen38-flash-next-v2-abliterated.hgn --revision e7783c5822dfda2785a0f5b435abf54df89fe8ee --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_CHECKPOINT=/v2/qwen38-flash-next-v2-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-v2-abliterated \
  -v /models/w4b:/models:ro -v /models/v2:/v2:ro \
  ghcr.io/peonist-ai/halogen-flash-server:0.15.0

• 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.

Then send a request:

curl -s http://127.0.0.1:8731/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"halogen-v2-abliterated","messages":[{"role":"user","content":"What is 37*43?"}]}'

GET /health reports the model, the engine version, the context settings and whether vision is enabled. The server repository documents every setting.

Benchmarks

Historical r1/r2 section; not validation of the current native conversion.

Quality

Scored with the engine's own ppl mode (teacher-forced, prefill kernels, 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, r2 beats r1 by 0.027 nats (t = −12.5, 95% CI −0.031 to −0.023; better on 41,719 tokens, worse on 29,035), and the advantage holds in every context-length band from 1k to 70k tokens.

The rounding was calibrated on the WikiText-2 train split. The test split is different text but the same kind of writing, which likely explains why r2 lands slightly below stock. Expect the gain over r1 to carry over to other text, but chat and code were not scored.

Speed

Test setup:

• Hardware: Ryzen AI Max+ 395 (Radeon 8060S), 128 GB unified memory.
• Server: halogen-flash-server 0.15.0, 262,144-position KV pool, 2 slots, HALOGEN_MAX_TOK=8192, MTP drafter.
• Requests: one at a time. Greedy means temperature 0; sampled means temperature 1 with a fixed seed. Prefill is a cold 34k-token prompt.

The r1 column shows two sessions. The second (after the slash) was measured back-to-back with r2 on a busier machine, so compare r2 against that one; the stock and GGUF numbers come from the first session. The differences are within run-to-run noise: r2 is no slower than stock.

The GGUF column is an earlier abliterated Qwen3.8-Flash-Next fine-tune, quantized to IQ4_XS and served by the same engine. It shows what running an abliterated GGUF costs compared with an abliterated .hgn. The GGUF is not published.

Speed in real use

These are the server's /metrics counters from r1 after 81 requests of mixed production traffic (chat, agentic tool use, greedy and sampled). r2 has the same draft head:

The server reports MTP and prompt-lookup drafts in one counter, so the acceptance rate cannot be split between them.

Refusal removal

Refusal-direction component of the writer tensors, |rᵀW|/|W|. Here r is the refusal direction, taken as the top singular vector of orcarouter's weights minus the base weights:

The 4-bit residual is consistent with 4-bit quantization noise.

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

Capability checks

These are smoke tests, one run each; r1 and r2 both passed:

No task benchmark (MMLU, GSM8K and so on) has been run; perplexity is the only quality measure.

Re-encoding accuracy

Relative L2 error of each re-encoded writer against orcarouter's BF16, over all 149 writers:

Calibrated rounding deliberately accepts slightly higher weight error in exchange for lower error in the layer's output on real activations. That is why the I4R weight error is higher than r1's round-to-nearest (0.097), while perplexity is better.

What changed

Historical r1/r2 section; not validation of the current native conversion.

orcarouter's BF16 model differs from base Qwen3.8-Flash-Next in exactly 149 residual-writer tensors. Scanning all 1,658 tensors confirmed this. The build takes those 149 tensors from orcarouter's BF16 and encodes each into the store v2 uses for that tensor:

Every other payload, including the MTP head's remaining tensors, is peonist's v2 byte for byte.

Calibration: input second moments (Hessians) for every changed HT and I4R tensor were collected from the base model's activations over 128 × 2,048 tokens of WikiText-2 train, layer by layer on CPU. The tokens and the manifest are in calibration/. This is not peonist's private calibration set, so the codes do not match peonist's exactly.

Reproduce

Historical r1/r2 section; not validation of the current native conversion.

tooling/README.md lists every input with its pinned revision, the disk, RAM and time needed, and notes on the reverse-engineered v2 store formats. In outline:

# 1. Calibrate once: ~100 min on CPU, ~24 GiB RAM, ~50 GB of Hessians.
python3 tooling/hgncalibrate.py --base base-qwen38-flash-next/ --out hessians/ \
  --sequences 128 --length 2048 --batch-size 8 --threads 8

# 2. Build: ~67 min with 8 workers. Needs the C++ HT helper (see tooling/README.md).
python3 tooling/halogen_v2_abliterate.py build \
  --v2 qwen38-flash-next-v2.hgn --donor orcarouter-bf16/ \
  --writers contract --ht-encode trellis --rounding ldlq --hessians hessians/ \
  --out qwen38-flash-next-v2-abliterated.hgn --workers 8

Compare the tokens_sha256 in your Hessian manifest with calibration/manifest.json to confirm the same calibration input. A byte-identical rebuild of r2 from a fresh calibration has not been tested; floating-point summation order may make it differ slightly.

The r1 build (--codebook-from qwen38-flash-next-w4b.hgn, no Hessians) is still supported and rebuilds byte-identically. The same tool can re-apply the donor's edit to a future v2 checkpoint, as long as the store layout stays compatible. It refuses to run on formats it doesn't recognize.

Limitations

Historical r1/r2 section; not validation of the current native conversion.

• Calibration used public WikiText-2 text, not peonist's calibration set. Quality was measured on the same domain; chat, code and long-context retrieval were not scored.
• 8,309 of 24,576 routed experts saw fewer than 1,024 calibration tokens (187 saw none) and fall back toward a layer-pooled Hessian.
• The four MTP/embedding writers in q4c/q8g64 use round-to-nearest.
• Refusal removal was tested on six prompts. It is not guaranteed for every prompt, and the model may still refuse some requests.
• The file only works with the Halogen engine. It was tested on 0.15.0 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 license and with the laws that apply to you.
• The uploader accepts no liability for misuse.

License and attribution

For current model weights, see the pinned source's full Qwen Community License 1.0 notice in LICENSE-QWEN. The original card's Apache-2.0 attribution below is retained as historical context, not as a replacement for that source notice. The existing Apache tooling licenses remain unchanged.

Historical r1/r2 attribution:

• Qwen/Qwen3.8-Flash-Next: the base model.
• peonist-ai/halogen-qwen3.8-flash-next: the v2 checkpoint. Its unmodified tensors are redistributed here.
• orcarouter/Qwen3.8-Flash-Next-Uncensored: the abliteration writer values.

Historical r1/r2 modification notice: the 149 tensors listed above were replaced with values re-quantized from orcarouter's weights. The calibration tokens are from WikiText-2 (CC BY-SA 3.0). "halogen" is a trademark of Peonist, LLC. The name is used here only to identify the compatible engine and file format. 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.