Qwen3.8-Flash-Next-Uncensored-NVFP4

Abliterated Flash-Next, packed so it serves on one GB10 / 96 GB Blackwell the same way Primitive's NVFP4 does.

NVFP4 graft of orcarouter/Qwen3.8-Flash-Next-Uncensored onto the Primitive experts-only RTN recipe. Routed experts are NVFP4 (group 16); the 51B PLE n-gram table stays BF16 so VLLM_PLE_CPU_OFFLOAD works. This is not the 336 GB BF16 source and not a TP=8 serve.

---

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream), inherited from the OrcaRouter BF16 parent. As a direct consequence:

• It will comply with harmful, unethical, offensive, or illegal requests that the original Qwen3.8-Flash-Next would refuse. It has no meaningful built-in guardrails.
• It is released strictly for legitimate research — interpretability, AI-safety and refusal-mechanism study, red-teaming, robustness evaluation, and controlled experiments.
• You assume full responsibility and liability for how you use it and for everything it generates. Do not deploy it to end users or in production without adding your own safety, moderation, and abuse-prevention layers.
• Use must comply with the Qwen Community License 1.0 inherited from Qwen/Qwen3.8-Flash-Next, and all laws and regulations that apply to you. (OrcaRouter labeled the BF16 parent Apache 2.0; the Qwen source card is Community 1.0, which is what this checkpoint follows.)
• The authors and uploaders accept no liability for any misuse or harm. Its outputs do not reflect the views of the uploaders, OrcaRouter, Primitive, or Qwen / Alibaba.

By downloading or using this model you acknowledge and accept the above.

---

What this is

Abliteration (from the OrcaRouter card)

Refusal-direction removal following Arditi et al. (2024). Residual-writing matrices only (149 tensors):

Never touched: vision + video tower, MoE router, fused experts.gate_up_proj, Hyper-Connection mixers, QSA indexer, n-gram embedding table, mtp.fc_*, norms, lm_head.

Abliteration is a weight edit, not data-level unlearning. Fine-tuning on refusal-heavy / safety data can partially re-introduce refusals.

Intended use

• Research into refusal mechanisms, alignment, and interpretability.
• Red-teaming and safety / robustness evaluation in controlled environments.
• A single-GPU NVFP4 serve of the uncensored Flash-Next weights.

OrcaRouter's BF16 eval (rule-based refusal classifier, not an LLM judge) is the capability / refusal reference for the parent; this graft does not re-run that suite. Sanity-checked on this checkpoint: text pong, vision OCR of a solid-red PNG, tool-calling parser flags identical to Primitive.

---

Serve with vLLM

Stock image: vllm/vllm-openai:qwen38-flash-next. Two flags are load-bearing on one GPU:

• --distributed-executor-backend mp — without this the n-gram offload worker never starts; the server looks healthy and hangs on the first forward.
• VLLM_PLE_OFFLOAD_READY_TIMEOUT=1800 — the worker loads a ~95 GB BF16 table; the 600 s default can expire first.

Do not use an FP8 n-gram table (RadixArk / Inferact-style). VLLM_PLE_CPU_OFFLOAD cannot load those.

Tool calls use Qwen3-Coder XML (--tool-call-parser qwen3_coder). Toggle thinking per request with chat_template_kwargs={"enable_thinking": true|false}. Pass image_url content parts for vision.

NVIDIA GB10 / DGX Spark (this checkpoint's home)

Unified memory is 121 GiB. Official “1× 96 GB GPU + ~100 GiB host RAM” is two pools; on GB10 they are the same pool, so the PLE table must page to a ~100 GiB swapfile or the load OOMs. --gpu-memory-utilization 0.75 (not 0.92). Hermes needs context ≥ 64K, so --max-model-len 65536.

# once per boot, before vLLM:
#   fallocate / swapon a ~100 GiB file (e.g. qwen38-ple.img)

docker run --name vllm --ipc=host --gpus all -p 8000:8000 \
  -e VLLM_PLE_CPU_OFFLOAD=1 \
  -e VLLM_PLE_OFFLOAD_READY_TIMEOUT=1800 \
  -e FLASHINFER_DISABLE_VERSION_CHECK=1 \
  -e CUTE_DSL_ARCH=sm_121a \
  -v /path/to/Qwen3.8-Flash-Next-Uncensored-NVFP4:/model \
  vllm/vllm-openai:qwen38-flash-next \
  /model \
  --served-model-name Qwen3.8-Flash-Next-Uncensored-NVFP4 \
  --host 0.0.0.0 --port 8000 \
  --trust-remote-code \
  --distributed-executor-backend mp \
  --gpu-memory-utilization 0.75 \
  --max-model-len 65536 \
  --max-num-seqs 4 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3

Discrete 96 GB Blackwell (Primitive's recipe)

Same image and PLE flags; raise utilization if GPU RAM and host RAM are separate:

docker run --gpus all --ipc=host -p 8000:8000 \
  -e VLLM_PLE_CPU_OFFLOAD=1 -e VLLM_PLE_OFFLOAD_READY_TIMEOUT=1800 \
  vllm/vllm-openai:qwen38-flash-next \
  --model mazinb/Qwen3.8-Flash-Next-Uncensored-NVFP4 \
  --distributed-executor-backend mp \
  --gpu-memory-utilization 0.92 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3

Host still needs ~100 GB free RAM (or swap) for the n-gram table.

---

Measured on DGX Spark (GB10)

Host edgexpert-3b5a, 121 GiB unified memory, vllm/vllm-openai:qwen38-flash-next, PLE CPU offload, --gpu-memory-utilization 0.75, --max-model-len 65536, --max-num-seqs 4, ~100 GiB PLE swap. Same NVFP4 layout as Primitive; decode speed is the recipe, not the abliteration overlay.

Closed-loop token sweep (thinking off)

ignore_eos=true, max_tokens=256, unique prefixes, 12 requests/level after 2 warmup. Closed-loop harness on this Spark, 2026-08-27, while Primitive NVFP4 was loaded (identical expert packing / PLE / vLLM flags).

c=1 is ~17 tok/s / ~58 ms per output token. Do not compare to Primitive's 74 tok/s on a discrete RTX PRO 6000 — that card has a separate 96 GB GPU + host RAM. On GB10 the PLE table lives in NVMe-backed swap.

What's quantized to what

MTP fused experts.down_proj stays BF16 (Primitive layout) and is overlaid from the abliterated source.

---

Bias, risks, and limitations

• Safety guardrails removed — the model will produce harmful, biased, or offensive content on request (see the disclaimer).
• It inherits any biases and limitations of Qwen3.8-Flash-Next and of NVFP4 RTN (no calibration).
• OrcaRouter's reported refusal metric is a rule-based heuristic on the BF16 parent; this NVFP4 graft was not re-scored on AdvBench / HarmBench.
• On GB10, PLE paging to NVMe makes decode slower than a discrete 96 GB GPU with 100 GiB of spare host RAM.

License

Qwen Community License 1.0, inherited from Qwen/Qwen3.8-Flash-Next. Abliteration and NVFP4 packing do not change those obligations.

Credit: Qwen / Alibaba (base), OrcaRouter (abliteration), Primitive (single-GPU NVFP4 + BF16 PLE recipe).