license: other
license_name: qwen-community-license-1.0
license_link: LICENSE
base_model: bottlecapai/ThinkingCap-Qwen3.8-Flash-Next
base_model_relation: quantized
pipeline_tag: text-generation
library_name: halogen
inference: false
tags:
- halogen
- halogen-flash
- qwen
- qwen3.8
- flash-next
- thinkingcap
- abliterated
- moe
- mixture-of-experts
- quantized
- hgn
- bf16-to-hgn
- token-efficient
- efficient-thinking
ThinkingCap Qwen3.8-Flash-Next — abliterated Halogen HGN (v2)
Native Halogen .hgn weights for
halogen-flash-server,
built frombottlecapai/ThinkingCap-Qwen3.8-Flash-Next
BF16 after a transferred-direction abliteration on the 149 residual-writer
contract (same method Quat3rnion used for Swift 1.5), then converted withbf16-to-hgn profile v2 (calibration-free).
These weights are Halogen’s .hgn format and will not load in Transformers,
vLLM, llama.cpp, or similar. Mount them into a halogen-flash-server
container (or equivalent Halogen runtime).
Hub upload:johnlockejrr/ThinkingCap-Qwen3.8-Flash-Next-abliterated-HGN.
Sibling packs:
| Pack | Hub |
|---|---|
| ThinkingCap v2 (no abliteration) | johnlockejrr/ThinkingCap-Qwen3.8-Flash-Next-HGN |
| ThinkingCap calibrated RCO | johnlockejrr/ThinkingCap-Qwen3.8-Flash-Next-calibrated-rco-HGN |
What was done
Direction transfer (BF16, CPU/NumPy, no GPU load)
Quat3rnion’s audited refusal directionr(2560-d, from stock Qwen vsorcarouter/Qwen3.8-Flash-Next-Uncensored,
published under his Swift abliterationabliteration/manifest.json)
was projected onto ThinkingCap’s 149 residual writers (attn outs / expert
downs / shared downs / MTP writers / PLE value / embeds). Non-writers were
copied unchanged. Tooling:thinkingcap_bf16_abliterate.py(wrapper around hisswift_bf16_abliterate.py
contract). Output BF16:/opt/halogen/ThinkingCap-Qwen3.8-Flash-Next-abliterated-bf16withabliteration-manifest.json(status: complete, gates passed).Halogen convert (no calibrated GPTQ)
That abliterated BF16 → HGN withbf16-to-hgnCUDA,--profile v2, policybf16-hgn-v2-rtn-nf4-signed-rms-trellis-v1. Abliteration already moves
weights; we did not add--quality calibrated(plain ThinkingCap HGN
already beat calibrated on our tool-eval board).
This is not a full heretic Optuna run on ThinkingCap (that needs a large
VRAM box). It is the same-arch direction implant Quat3rnion validated for
Swift-style finetunes — better than raw orca tensor swap, not as native as
heretic-on-ThinkingCap.
Upstream model (behavior)
ThinkingCap is BottleCap AI’s shorter-thinking post-train of Flash-Next.
Upstream reports ~38.5% fewer thinking tokens on average at ~88.8% vs
~89.8% base accuracy:
ThinkingCap card
and blog.
Abliteration may change refusal behavior and can cost some capability. Treat
refusal removal as empirical / unverified until you re-bench.
What is here
| File | Size | Role |
|---|---|---|
model.hgn |
62.11 GiB | Trunk + embedded MTP (qwen3.8-flash-next-v2) |
qwen38-flash-next-ngram.hgn |
47.68 GiB | N-gram table (required) |
qwen38-flash-next-vision.hgn |
0.84 GiB | Vision tower (optional; HALOGEN_VISION_TOWER=1) |
tokenizer/ |
— | Tokenizer + chat template (from ThinkingCap) |
config.json |
— | Model config |
conversion-manifest.json |
— | Convert provenance / plan |
sha256.txt |
— | SHA-256 of the three .hgn payloads |
cap_header.png / thinkingcap_demo.mp4 |
— | Upstream branding / demo |
Rough total: ~111 GiB. No separate mtp.hgn. N-gram / vision hashes
match our other Flash-Next v2 packs; only model.hgn is unique to this
abliterated ThinkingCap convert.
How it was made
BF16 edit: bottlecapai/ThinkingCap-Qwen3.8-Flash-Next
+ Quat3rnion audited r (Swift abliteration manifest)
→ ThinkingCap-Qwen3.8-Flash-Next-abliterated-bf16
method: transferred-direction abliteration (ThinkingCap target)
writers: 149-contract (not full heretic)
HGN convert:
source: ThinkingCap-Qwen3.8-Flash-Next-abliterated-bf16 (HF safetensors, 131 shards)
tool: bf16-to-hgn 0.1.0 (CUDA, HGN_CUDA_ARCH=86)
profile: v2
policy: bf16-hgn-v2-rtn-nf4-signed-rms-trellis-v1
quality: (default / not calibrated)
hgn-spec: 1.1.0 @ 79e75fd4ae7642478823525b0a3212951f584af5
# 1) BF16 direction transfer (CPU)
cd /opt/halogen/tooling-swift-abl/tooling
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 \
/opt/halogen/.venv/bin/python thinkingcap_bf16_abliterate.py \
--target /path/to/ThinkingCap-Qwen3.8-Flash-Next \
--direction /path/to/abliteration/manifest.json \
--output /path/to/ThinkingCap-Qwen3.8-Flash-Next-abliterated-bf16
# 2) Convert + verify (no --quality calibrated)
BIN=/path/to/bf16-to-hgn-optimal/target-cuda/release/bf16-to-hgn
SRC=/path/to/ThinkingCap-Qwen3.8-Flash-Next-abliterated-bf16
OUT=/path/to/ThinkingCap-Qwen3.8-Flash-Next-abliterated-HGN
$BIN convert --source "$SRC" --output "$OUT" \
--profile v2 --seed 0 --threads 8 --ht-beam 128 \
--ht-backend cuda --i4r-backend cuda --q4cp-backend cuda
$BIN verify --source "$SRC" --output "$OUT" --decode-backend cuda \
> verify-thinkingcap-abliterated-r-transfer.json
This build: convert published successfully; verifyprovenance_attested: true (1658 source / 2349 output tensors,
118 785 459 868 payload bytes).
HF inputs do not take --mtp / --vision / --ngram / --assets.
How to run
hf download johnlockejrr/ThinkingCap-Qwen3.8-Flash-Next-abliterated-HGN \
--local-dir ~/halogen-models/ThinkingCap-Qwen3.8-Flash-Next-abliterated-HGN
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/ThinkingCap-Qwen3.8-Flash-Next-abliterated-HGN:/models:ro \
ghcr.io/peonist-ai/halogen-flash-server:0.19.1
On Docker, use --group-add video --group-add render. API on :8731.
Server docs: halogen-flash-server.
For greedy tool-eval-bench against Halogen, prefer --parallel 1.
Precision (v2)
| Checkpoint | ~62.1 GiB, Halogen v2 policy |
| Routed experts | I4R |
| Dense HT / LM head | HT4 |
| Hyper-connection mixers | Q6G64 |
| Embeddings / small projections | Q4C-P |
| MTP dense / mixer | Q8G64 (in model.hgn) |
| Vision | BF16 sidecar |
| N-gram | FP8G, own file |
Credits
- BottleCap AI —
bottlecapai/ThinkingCap-Qwen3.8-Flash-Next. - Qwen / Alibaba — Flash-Next base; license in
LICENSE. - OrcaRouter —
donor used to audit the refusal directionr(not copied as tensor transplant). - Quat3rnion —
swift_bf16_abliterate.py
/ auditedabliteration/manifest.json, andbf16-to-hgn. - peonist-ai —
halogen-flash-server.
SHA-256
85aa001000712cb5191a6eae1ce00bcab94989df2a56423e3fd755a061581b66 model.hgn
339f5c508ca592ee2613ac5c3082587a88991d5592c4160d49a32a6dbdb3f30b qwen38-flash-next-ngram.hgn
b6ca6716eaab601773f7155701bb09f39c72922caa21620bab577f0281f2ea2a qwen38-flash-next-vision.hgn
N-gram / vision match sibling Flash-Next v2 packs; only model.hgn is unique to this abliterated convert.
Disclaimer
Community Halogen conversion of direction-abliterated ThinkingCap BF16.
Not an official BottleCap, Qwen, OrcaRouter, Quat3rnion, or peonist release.
Refusal removal and quality are not guaranteed — re-measure on Halogen
(tool-eval, your own probes) before production use. Provide moderation if
you deploy to others.
