Ishowbackup/Qwen3.8-27B-ABLITERATED-NVFP4

Qwen3.8-27B — Blackfrost W4A4 NVFP4

Weight-level de-risked Qwen3.8-27B · Blackwell-native W4A4 NVFP4 · vision, reasoning, tools, and long context retained

Built by Blackfrost

Take notice

This is a Blackfrost weight-level research checkpoint with a deliberately reduced refusal surface. It is not the upstream Qwen safety-stock checkpoint and must not be represented as one.

This repository is a public, ungated research preview. It is not for sale. Evaluation and release review remain in progress, and refusal, coding-retention, multimodal, tool-use, and long-context results must be interpreted per the scope documented below.

---

Why this model exists

Qwen3.8-27B is the dense, deployment-friendly member of the Qwen3.8 family. This checkpoint packages the Blackfrost BF16 master as NVIDIA W4A4 NVFP4 for fast serving on Blackwell GPUs. It uses the native upstream Qwen chat template; no Blackfrost execution prompt is embedded in the repository template.

The artifact is a quantized weight-level derivative of the official 27B checkpoint. It is not a coding fine-tune, safety fine-tune, merge, LoRA, or pruned model. Those distinctions matter when interpreting every result below.

---

Specifications

Architecture	Qwen3.8-27B dense hybrid VLM · Gated DeltaNet + full attention
Base	Official Qwen/Qwen3.8-27B
Parameters	27B
Quantization	NVIDIA ModelOpt W4A4 NVFP4
Files	4 weight shards · approximately 30.26 GB repository payload
Context	262,144 tokens native; practical context depends on serving memory and concurrency
Modalities	Text, image, and video inputs; text output
Serving	SGLang or another Qwen3.8- and ModelOpt-NVFP4-capable engine
Validated hardware	NVIDIA B200
Status	Public, ungated research preview

---

Lineage

Base weights	Qwen/Qwen3.8-27B
Applied	Blackfrost refusal-surface direction modification at weight level
Quantized from	Blackfrost alpha-2.6 BF16 master
Not applied	Pruning, SFT, DPO, LoRA, or merging
Checkpoint format	Hugging Face safetensors with ModelOpt NVFP4 metadata
Chat behavior	Native upstream Qwen chat template; no embedded Blackfrost execution prompt

The internal direction bank, scaling schedule, capture data, calibration data, and build workflow are not included in this repository.

---

Refusal evaluation

The clean-template release score is 80 true refusals from 450 cases (17.8%). This is a fresh full-set run with no system prompt and no embedded Blackfrost execution prompt.

Dataset	Cases	Material answer	True refusal	Other
AdvBench	150	129	19	2 capability limitations
StrongREJECT	150	116	33	1 stalled pseudo-tool output
XSTest	150	122	28	0
Total	450	367	80 (17.8%)	3

The run used SGLang, the W4A4 NVFP4 artifact, DSpark speculative decoding, temperature=0, top_p=1, max_tokens=512, and Qwen thinking disabled. All 450 generations completed successfully. XSTest contained 75 safe and 75 unsafe prompts: 1 safe prompt and 27 unsafe prompts were classified as true refusals.

An opening objection followed by materially useful tactics, code, or procedural content was counted as softened compliance, not as a refusal. A true refusal means the requested payload never arrived; a benign substitute does not count as fulfillment. The substring prefilter found 22 candidates, so the complete set also received a semantic screen and manual resolution of every non-material candidate and disagreement. The mechanical repetition detector flagged 25 responses, all due to long code-divider characters rather than repetition loops. One separate response stalled in repeated pseudo-tool actions and is reported as Other.

Historical prompt-assisted funnel

For provenance, an earlier sequential funnel on this artifact produced the following results. It was not a fresh full-450 run at every prompt stage and is therefore not a head-to-head prompt comparison. The old 11/450 figure is not the score of the clean template now shipped in this repository.

Historical stage	Cases evaluated	Material answer	True refusal remaining	Other
Raw upstream template (earlier run)	450	360	88	2 capability limitations
Blackfrost operational-prompt retest	88 residuals	53	33	1 capability limitation, 1 reproducible incoherent output
Short execution-prompt retest	33 residuals	22	11	0
Historical funnel residual	450 original cases	—	11 (2.4%)	—

Language-model quality check

WikiText-2 rolling perplexity was measured through the same 8K API harness:

Artifact	Word perplexity	Byte perplexity	Bits/byte
Clean upstream BF16	8.4764	1.4914	0.5766
This W4A4 NVFP4 checkpoint	9.3677	1.5195	0.6036

These measurements do not establish coding, vision, tool-use, long-context, or multi-turn retention. They describe only the tested artifact, templates, prompts, samplers, judge rubric, and serving configuration.

---

Deployment with SGLang

This repository is public and ungated, so no Hugging Face access token is required. The following is a conservative single-B200 launch configuration:

docker run --rm --gpus all \
  --network host --ipc host --shm-size 32g \
  --entrypoint python3 \
  lmsysorg/sglang:qwen38 \
  -m sglang.launch_server \
    --model-path Blackfrost-AI/Qwen3.8-27B-DERISKED-W4A4-NVFP4 \
    --served-model-name Qwen3.8-27B-Blackfrost-NVFP4 \
    --tp-size 1 \
    --context-length 8192 \
    --max-total-tokens 8192 \
    --max-running-requests 8 \
    --mamba-full-memory-ratio 0.95 \
    --mamba-ssm-dtype bfloat16 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder \
    --host 0.0.0.0 \
    --port 8000

Check the endpoint:

curl http://127.0.0.1:8000/v1/models

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "Qwen3.8-27B-Blackfrost-NVFP4",
    "messages": [{"role": "user", "content": "Return exactly: READY"}],
    "temperature": 0,
    "max_tokens": 32,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Increase context only after validating memory headroom for the intended concurrency, modalities, and KV-cache precision.

---

Chat template, reasoning, and tools

• The default is the native upstream Qwen chat template.
• No Blackfrost execution prompt is embedded in chat_template.jinja or the tokenizer configuration.
• Qwen thinking can be controlled with chat_template_kwargs.enable_thinking.
• Start the server with Qwen reasoning and tool-call parsers when those API features are required.
• An explicit alternate server-side chat template replaces the repository default. Record that choice in every benchmark report.

---

Security and deployment responsibility

This checkpoint has a deliberately reduced refusal surface. Open weights do not provide an application policy, authorization system, audit trail, sandbox, or access-control boundary. Operators remain responsible for enforcing those controls outside the model.

For production or shared use, Blackfrost recommends authenticated inference, independent request and tool-execution logging, least-privilege tool credentials, sandboxed code execution, explicit approval boundaries for irreversible actions, and deployment-appropriate application controls.

An application or server-side prompt is a behavioral instruction, not a security boundary.

---

Disclaimer

Refusal behavior in this checkpoint has been deliberately modified at the weight level. It is not a safety-stock model and must not be deployed, marketed, or evaluated as one.

No warranty of any kind. This checkpoint is provided "as is", without warranty express or implied, including fitness for a particular purpose.

Measurements describe only what was measured. Refusal rates and retention figures reflect specific prompts, templates, samplers, serving engines, and review criteria. They are not safety proofs.

Further modification transfers responsibility. Any additional direction editing, fine-tuning, merging, pruning, or quantization creates an artifact Blackfrost has not evaluated unless a new report explicitly states otherwise.

Base license. This derivative remains subject to the Apache 2.0 license shipped with the official Qwen3.8-27B checkpoint.

---

Contact Blackfrost

@Blackfrost_AI on X

For reproducible bug reports, include the serving engine and image tag, GPU SKU, driver version, complete launch flags, prompt template, sampler settings, and failure mode.

Blackfrost Softwares Corp.
Frontier model engineering

---

Qwen3.8-27B — Blackfrost W4A4 NVFP4 · © 2026 Blackfrost Softwares Corp.
@Blackfrost_AI