Blackfrost-AI/PINQWEN-3.5-9B-1M-NVFP4

PINQWEN‑3.5‑9B‑1M — NVFP4

🪟 A 9B that reads a million tokens, thinks before it speaks, and never refuses — now in 4‑bit for Blackwell.

PINQWEN‑3.5‑9B‑1M is Blackfrost AI's compact powerhouse — a reasoning model distilled through The Void, our multi‑teacher reasoning‑distillation method, on the Qwen 3.5 9B architecture. It reasons first and answers second (<think> → answer), across a full one‑million‑token context window. This is the NVFP4 build: 4‑bit weights tuned for blazing‑fast serving on NVIDIA Blackwell.

---

🚀 Why PINQWEN‑3.5‑9B‑1M?

🪟 1,000,000‑token context	Whole codebases and document sets in a single prompt — YaRN‑extended on a hybrid backbone built for fast long context.
🔓 Uncensored	Answers directly, no reflexive refusals. You're in control.
🧠 Reasons in the open	Native <think>…</think> chain‑of‑thought from a panel of frontier teachers.
⚡ Blackwell‑native	Weight‑only NVFP4 (group‑size 16) — a fraction of BF16's footprint, tuned for RTX PRO 6000 / B‑series.

📦 This repository — NVFP4

A weight‑only NVFP4 (4‑bit, group‑size 16) build of the text backbone for NVIDIA Blackwell, with attention and output layers kept at higher precision. This is the text serving variant (vision not included) — use the BF16 or GGUF repo for multimodal. Full 1M context.

⚡ Quickstart (vLLM on Blackwell)

vllm serve Blackfrost-AI/PINQWEN-3.5-9B-1M-NVFP4 \
  --quantization modelopt_fp4 --kv-cache-dtype fp8 \
  --reasoning-parser qwen3 --max-model-len 1048576

🎛 Formats — every one is 1M context

Format	Repository	Best for
BF16	PINQWEN-3.5-9B-1M-BF16	Reference precision, fine‑tuning, vision
NVFP4	PINQWEN-3.5-9B-1M-NVFP4	Fast serving on NVIDIA Blackwell
GGUF	PINQWEN-3.5-9B-1M-GGUF	llama.cpp / local — full quant ladder + MTP

📄 License

Apache‑2.0. Base architecture: Qwen 3.5 9B. As an uncensored model, use responsibly.

---

PINQWEN‑3.5‑9B‑1M — part of Blackfrost AI's Void model family. 🖤