license: other
license_name: lil-license-1.0
license_link: LICENSE
base_model: local-inference-lab/Qwen3.8-Flash-Next-FLARE
base_model_relation: finetune
pipeline_tag: image-text-to-text
tags:

• flare
• nvfp4
• mxfp8
• abliterated
• uncensored

Qwen3.8-Flash-Next-FLARE-Abliterated

This model is based on Qwen3.8-Flash-Next-FLARE by Local Inference Lab, Inc., a non-profit organization, available at https://huggingface.co/local-inference-lab/Qwen3.8-Flash-Next-FLARE. Qwen3.8-Flash-Next-FLARE is licensed under the Local Inference Lab License, Version 1.0.

Reduced-refusal model, and an incomplete edit. A refusal-direction edit was
applied on purpose, and it covers 24,576 of the 25,188 tensors the reference
build edits. Read "What is and is not edited" and "Integrity" before using it.

A prepared Flare runtime package — the gb202 revision, for a single NVIDIA
RTX PRO 6000 Blackwell — with abliterated routed-expert weights written into its
payload in place. Everything else in the package is the vendor's original bytes:
the compiled CUDA kernels, the captured execution graphs, the PLE n-gram table,
the tokenizer and the chat template.

Integrity — read this first

Run it without --verify-weights. The package carries a single digest over
its whole 77.78 GB weight region, and editing weights invalidates it. There is no
public tool that can re-sign a Flare package, so that digest has not been
regenerated.

Everything else loads normally, because the stored-data digest is only consulted
when you ask for it. This is a real regression against the original package: a
recipient cannot cryptographically verify these weights. Treat the table above as
the honest statement of what you lose.

What is and is not edited

Edited — 24,576 planes, every main-decoder routed-expert down_proj
(48 layers × 512 experts), NVFP4, written as byte-identical replacements of the
same length at the same offsets:

model.language_model.layers.{0..47}.mlp.experts.{0..511}.down_proj.weight

Not edited, though the reference build edits them — 612 tensors:

No scale tensor, no router, no gate_proj or up_proj, no q/k/v_proj, no
vision tensor and no PLE table entry is touched — not here and not in the
reference build.

Measured effect

Scored on 16 prompts from the public treadon/abliteration-eval harmful split,
protocol A — thinking disabled, which is the regime the direction was captured
in — with a 47-marker refusal scorer. Only refusal form was scored.

n is 16, so the interval is wide — roughly 4 % to 46 %. The result is
unambiguous about direction and weak about magnitude. Read it as "the edit took
effect and is incomplete", not as a rate you can quote.

The gap against the reference build is informative in its own right: routed
experts are 97.6 % of the edit by tensor count but clearly not 100 % of the
effect. The residual refusal is attributable to the 612 unedited tensors, which
are the attention, shared-expert and embedding write paths.

No capability evaluation was run on this package. For the same weights in a plain
checkpoint, with full measurements, see
huginnfork/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD-Abliterated,
which carries a complete 25,188-tensor edit and its evaluation files.

Serving

Requires Flare. These files cannot be loaded by Transformers, vLLM, SGLang or
llama.cpp.

A driver of at least CUDA 13.2 is required. Measured: driver 580.178.04 /
CUDA 13.0 fails during Prepare execution with
cuGraphKernelNodeSetAttribute failed (1, CUDA_ERROR_INVALID_VALUE), while
595.71.05 / CUDA 13.2 and 615.71.09 / CUDA 13.4 both work.

hf download huginnfork/Qwen3.8-Flash-Next-FLARE-Abliterated \
  --include '*.safetensors' --local-dir ./models

flare run ./models/qwen3.8-flash-next-tp1-c16-mtp5.safetensors \
  --gpus 0 --embedding-table-storage ram \
  --publisher-keys ./publisher.json \
  --trust-store ./flare-state/trust.toml

publisher.json is included here and is the vendor's own profile, copied
unchanged from the original repository's main branch — the weights branches do
not carry it. The signing key is the vendor's and the signature over the metadata
still verifies. For an unattended start, add:

  --dangerously-approve-key 937c0143f6856d3ee8ad164f62e3d804bd825d7a8750d0accb9a5849f703a8c1 \
  --dangerously-accept-publisher local-inference-lab

Prepared for one GPU, up to 16 concurrent requests and up to five MTP draft
tokens. --embedding-table-storage ram needs about 26.8 GiB of host RAM; disk
is the default and needs far less.

Flare rejects temperature=0 — it returns
greedy decoding incorrectly collapses the model's probability distribution.
Use 0.01 for near-deterministic sampling. Package defaults are temperature 1.0,
top-p 0.95, top-k 20.

Safety

Refusal behaviour has been deliberately reduced. This model answers requests the
original declines. Do not deploy it as a general-purpose assistant without your
own safeguards.

A refusal number is a protocol and scorer pair, and the protocol dominates. The
figure above is protocol A, thinking off, which is the harder regime — the
unedited parent refuses 100 % there against 89.7 % with thinking on. Do not
compare it against a number measured any other way.

Because the edit is incomplete, behaviour is less predictable than either the
original or a fully edited build. It refuses sometimes and not others.

Licence and credits

Local Inference Lab License, Version 1.0. The full text is in LICENSE, with
NOTICE and LICENSES/, copied from the location the original repository's
license_link points to. The original repository ships no licence file of its
own.

The upstream model is Qwen/Qwen3.8-Flash-Next at
de4b8e4d43b917e7706784d8bb445c9af86a3540, under the Qwen Community License
1.0, whose text is also included. Its condition 2 requires a separate licence
from Qwen for Model-as-a-Service or AI-assistant commercial use.

Redistribution here is by permission of the Licensor. Neither Local Inference Lab
nor Qwen endorses this model or its modification.

• Qwen/Qwen3.8-Flash-Next — base model, by Qwen.
• local-inference-lab/Qwen3.8-Flash-Next-FLARE — the prepared runtime package,
  its compiled kernels and captured graphs, by Local Inference Lab, Inc.
• The Flare runtime, the NVFP4-CSF format and the b12x kernels — by Local
  Inference Lab, Inc.