license: other
license_name: lil-license-1.0
license_link: LICENSE
library_name: flare
base_model:

• lflare/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD-Abliterated
• local-inference-lab/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD
  base_model_relation: finetune
  pipeline_tag: image-text-to-text
  tags:
• flare
• nvfp4
• mxfp8
• abliterated
• reduced-refusal
• yarn
• gb202

This model is based on Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD by Local Inference Lab, Inc., a non-profit organization, available at https://huggingface.co/local-inference-lab/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD. Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD is licensed under the Local Inference Lab License, Version 1.0.

Qwen3.8-Flash-Next-FLARE-Abliterated-YaRN-512k

Reduced-refusal model with a 524,288-token context. This is a Flare runtime
package of lflare's abliterated Qwen3.8-Flash-Next,
exported with YaRN factor 2, so one RTX PRO 6000 Blackwell serves 512k tokens
per request instead of 262k. Read the Safety section before you deploy it.

At a glance

What differs from lflare's package

lflare/Qwen3.8-Flash-Next-FLARE-Abliterated
is the reference package of the same edit. This one is a separate export:

• YaRN factor 2 → 524,288 tokens, via Ignition's --rope-parameters
  (pull request #212).
• Concurrency 8 instead of 16. The longer context needs the KV room.
• --force-moe-decode-a16 at export.
• Newer Flare tree: upstream main 0de3041c plus PR #212 15d8b0b0.
• Quantized vision tower. The abliterated checkpoint keeps vision in BF16,
  and Ignition requires it quantized, so the 499 quantized vision tensors were
  taken byte-identical from Local Inference Lab's container. Text weights are
  untouched. Details in export.json.

Download and run

hf download Dsturb/Qwen3.8-Flash-Next-FLARE-Abliterated-YaRN-512k \
  --revision v1 --include 'model.safetensors' 'weights/*' 'deployments/*' \
  'publisher.json' 'config.json' --local-dir ./pkg

mkdir -p flare-state
docker run --rm -it --gpus 'device=0,capabilities=compute' --ulimit memlock=-1:-1 \
  -p 127.0.0.1:8000:8000 \
  --mount "type=bind,src=$PWD/pkg,dst=/pkg,readonly" \
  --mount "type=bind,src=$PWD/flare-state,dst=/state" \
  ghcr.io/local-inference-lab/flare:latest \
  run /pkg --gpus 0 --port 8000 --verify-weights \
  --embedding-table-storage ram --trust-store /state/trust.toml \
  --publisher-keys /pkg/publisher.json \
  --dangerously-accept-publisher local-development \
  --dangerously-approve-key 7ab239ecbc30aa5039823aa9ad42daabe0e221b4bcb56ed50e9aaeff01e9e824

config.json is a download-stats anchor, not a model config — the same
approach as lflare's repository. It deliberately has no model_type, so
AutoConfig.from_pretrained is expected to fail on it.

Measured

All on one RTX PRO 6000 Blackwell at 300 W, measured with
lil-bench 1.7.1 (standard profile,
isolated, no foreign traffic):

Prefill: 11,745 tok/s at 32k and 11,629 tok/s at 128k (same run).

Long context: a 446,870-token prompt with a needle in the middle was answered
correctly in 43.5 s. On load the runtime reports 524,288 tokens per request and
a KV cache of 1,318,912 tokens.

Refusal and capability (lflare's measurements, same weights)

The text weights are lflare's rank-2 dual-regime edit, unchanged; only the
context scaling, the export capacities and the packaging differ. The figures
below are lflare's, measured on the plain checkpoint
lflare/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD-Abliterated
(raw files in its measurements/), and were not re-run on this package.

Held-out rows, 136 harmful and 59 over-refusal, parent against the edit:

Capability in the same session: lavd 30/30 against the parent's 30/30,
estonia 30/30 against 30/30, paired GPQA-diamond 166/198 against 164/198
(not significant, p = 0.85).

How it was built

Ignition export on 2026-11-08, single GPU, about 45 minutes. The exact command,
the source revisions, both checkpoint changes and the signer are in
export.json. No sampling parameter was set or changed; the package ships with
the defaults Ignition derived.

Verifying the download

flare run ./pkg --verify-weights --stage-only \
  --publisher-keys ./pkg/publisher.json \
  --dangerously-accept-publisher local-development \
  --dangerously-approve-key 7ab239ecbc30aa5039823aa9ad42daabe0e221b4bcb56ed50e9aaeff01e9e824

This checks every payload object against the signed manifest and exits before
the GPU is used. SHA256SUMS lists every file. The primary:

file   : model.safetensors   (revision v1 = gb202 = main)
size   : 15,078,160 bytes
sha256 : be917c99310e774299d1aad33732271186e24bf11e4746f663ac2fbbf40afcd9

Limitations

• Concurrency 8, not 16. Aggregate C8 throughput is about 10 % below lflare's
  c16 package at the same power limit.
• Refusal behaviour, capability and vision were not re-measured through Flare
  on this package; the refusal figures above are lflare's.
• No KLD or perplexity comparison against the native 262k context exists for
  this model at any YaRN factor. Factor 2 was chosen over factor 4 (1M) to stay
  closer to the native positions; the 446,870-token needle test is the only
  long-context check here.
• Signed by a development key; the runtime announces it as "not an official
  published checkpoint". Approving the key is your trust decision.
• One GPU, TP1, RTX PRO 6000 Blackwell (gb202) only.
• Requires Flare. These files cannot be loaded by Transformers, vLLM, SGLang or
  llama.cpp.

Safety

Refusal behaviour has been deliberately removed by the upstream edit. The model
answers requests the parent declines, including harmful ones. Do not deploy it
as a general-purpose assistant without your own safeguards.

Why this exists: I use it to test my own security infrastructure, where
aligned models often stop before the task is done.

Lineage and credits

• Qwen/Qwen3.8-Flash-Next by Qwen.
• NVFP4/MXFP8 QAD container, Flare runtime, b12x kernels and Ignition by
  Local Inference Lab, Inc.
• Refusal edit by lflare.
• YaRN 512k export and packaging by Dsturb.

License

Local Inference Lab License, Version 1.0. See LICENSE, NOTICE and
LICENSES/. The Local Inference Lab texts are copied unchanged, with the
Identifying Marks intact.

Published with permission of the Local Inference Lab team under Section 2.4.
Local Inference Lab does not sponsor or endorse this model or its authors.
The upstream model remains under the Qwen Community License 1.0; see
LICENSES/LicenseRef-qwen-community-1.0.txt. If you operate a
Model-as-a-Service or AI-assistant business, that licence requires a separate
licence from Qwen before commercial use.