Qwen3.8-Flash-Next — Blackfrost NVFP4
Quantized Blackfrost Qwen3.8-Flash-Next · routed-expert NVFP4 · FP8 PLE
Built by Blackfrost
Weights published
This repository contains the Blackfrost NVFP4 checkpoint as 206 indexed safetensors shards with its verified configuration, tokenizer, and Qwentium chat template.
Blackfrost's proprietary modification process and associated research artifacts are not included in this repository.
Why this model exists
Qwen3.8-Flash-Next is Qwen's experimental preview of the architecture leading into Qwen4. It combines a large sparse model with low active compute, native multimodality, hybrid attention, four gated residual branches, n-gram embeddings, and multi-token prediction.
This checkpoint is the reduced-memory NVFP4 release of the Blackfrost BF16 research edition. It preserves the parent model's weight-level refusal-surface modification and Qwentium operational chat template while quantizing the routed-expert projections for efficient serving on supported NVIDIA hardware.
It is not a fine-tune, merge, LoRA, or pruned model.
Specifications
| Architecture | Qwen4ExpForConditionalGeneration · hybrid vision-language MoE |
| BF16 parent | Blackfrost-AI/Qwen3.8-Flash-Next-DERISKED-BF16 |
| Upstream | Qwen/Qwen3.8-Flash-Next on ModelScope |
| Language-model parameters | 125B total · 6B activated |
| Additional parameters | 51B n-gram embedding · 4B MTP |
| Layers / experts | 48 layers · 512 routed experts · 10 routed + 1 shared expert active per token |
| Attention layout | 36 Gated DeltaNet layers · 12 Qwen Sparse Attention layers |
| Gated residual | 4 branches · bottleneck rank 320 |
| Precision | Routed experts NVFP4 W4A4 · PLE tables FP8 · remaining weights BF16 |
| Indexed weight size | 135,195,303,851 bytes |
| Weight files | 206 indexed safetensors shards |
| Context | 262,144 tokens native; extended-context support follows upstream Qwen guidance |
| Modalities | Text, image, and video inputs; text output |
| Validated serving shape | 2× NVIDIA B200 · tensor parallel 2 |
| Status | Public NVFP4 release · weights available |
Lineage
| Base checkpoint | Blackfrost BF16 parent linked above |
| Quantization format | NVIDIA ModelOpt NVFP4 with FP8 PLE tables |
| Quantization reference | Compatible layout and calibration from RadixArk/Qwen3.8-Flash-Next-NVFP4 |
| Not applied | SFT, DPO, LoRA, merging, or expert pruning |
| Checkpoint format | Hugging Face-compatible safetensors with ModelOpt metadata |
| Chat behavior | Qwentium operational system prompt embedded in the default Qwen chat template |
Blackfrost's proprietary modification process and associated research artifacts are not included in this repository.
Artifact verification
The published checkpoint passed structural and serving verification before release:
- all 206 weight shards are indexed and readable;
- routed-expert NVFP4 values were rebuilt from the selected BF16 parent and passed deterministic requantization checks;
- non-target model state remains exact to the compatible quantized reference;
- configuration, tokenizer, processor, and Qwentium template files are complete; and
- the checkpoint loaded and completed deterministic generation on the validated serving shape.
These are artifact-integrity checks. They are not capability benchmarks or safety proofs.
Behavioral evaluation
No behavioral score is published for this quantized checkpoint until its evaluation has completed manual adjudication. Results from the BF16 parent must not be attributed to this NVFP4 artifact.
Qwentium chat template
The default Qwen chat template includes the Blackfrost Qwentium operational prompt. The template preserves Qwen's message structure, tool descriptions, multimodal inputs, reasoning controls, and caller-supplied system context.
The embedded prompt is a behavioral instruction, not an authorization system or security boundary. If a serving harness supplies a replacement chat template, it may bypass the embedded prompt. Record the exact template and template arguments in every evaluation report.
Deployment
The command below records the SGLang configuration validated for this published checkpoint.
docker run --rm --gpus all \
--network host --ipc host --shm-size 64g \
lmsysorg/sglang@sha256:12d3392bdc8be8d35e9a95f191df6aef99c5114bdbefd41bfdc7e760e6d25ec1 \
serve \
--model-path Blackfrost-AI/Qwen3.8-Flash-Next-DERISKED-NVFP4 \
--served-model-name Qwen3.8-Flash-Next-Blackfrost-NVFP4 \
--host 0.0.0.0 \
--port 8000 \
--tp 2 \
--quantization modelopt_fp4 \
--fp4-gemm-backend flashinfer_cutlass \
--page-size 64 \
--mamba-radix-cache-strategy extra_buffer \
--mamba-track-interval 64 \
--chunked-prefill-size 4096 \
--max-running-requests 36 \
--context-length 262144 \
--mem-fraction-static 0.80 \
--allow-auto-truncate \
--linear-attn-prefill-backend flashinfer \
--linear-attn-decode-backend flashinfer \
--mamba-ssm-dtype bfloat16 \
--reasoning-parser auto \
--sampling-defaults model
The model's recommended sampling defaults are temperature 1.0, top-p 0.95, and top-k 20.
OpenAI-compatible health and generation checks:
curl http://127.0.0.1:8000/health
curl http://127.0.0.1:8000/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{
"model": "Qwen3.8-Flash-Next-Blackfrost-NVFP4",
"messages": [{"role": "user", "content": "Return exactly: READY"}],
"temperature": 0,
"max_tokens": 64
}'
The validated configuration uses two B200 GPUs and BF16 KV cache. Increase context or concurrency only after measuring memory headroom for the actual workload, modalities, and cache settings.
Security and deployment responsibility
This checkpoint has a deliberately reduced refusal surface inherited from its BF16 parent. 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 endpoints, independent request and tool-execution logging, least-privilege credentials, sandboxed code and file access, and explicit human approval for irreversible actions.
License and disclaimer
This derivative remains subject to the Qwen Community License 1.0 shipped with the upstream Qwen3.8-Flash-Next release. Review the upstream license before commercial deployment, especially its separate-license condition for Model-as-a-Service and AI Work Assistant businesses.
Refusal behavior has been deliberately modified at the weight level in the BF16 parent. This is not a safety-stock checkpoint and must not be deployed, marketed, or evaluated as one.
No warranty of any kind. The checkpoint is provided "as is." Nothing here guarantees that any input will be accepted or refused, that every upstream capability is retained, or that any category of output is unreachable. Measurements describe only the exact artifacts, templates, samplers, serving engines, and review criteria under which they were produced.
Contact Blackfrost
@Blackfrost_AI on X
For reproducible bug reports, include the serving engine and image digest, GPU SKU, driver version, complete launch flags, prompt template, sampler settings, and observed failure mode.
Blackfrost Softwares Corp.
Frontier model engineering
Qwen3.8-Flash-Next — Blackfrost NVFP4 · © 2026 Blackfrost Softwares Corp.
@Blackfrost_AI
- Downloads last month
- 136