Blackfrost-AI/PINQWEN-3.5-9B-1M-BF16

PINQWEN‑3.5‑9B‑1M

🪟 A 9B that reads a million tokens, thinks before it speaks, and never refuses.

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: for any real problem it opens a <think> block, works it through, then delivers a clean, direct answer — and it does it across a full one‑million‑token context window that most models this size can only dream of.

---

🚀 Why PINQWEN‑3.5‑9B‑1M?

🪟 1,000,000‑token context	Feed it entire codebases, whole books, months of chat logs — at once. A YaRN‑extended window on a gated‑linear‑attention hybrid backbone built to make long context fast, not just possible.
🔓 Uncensored	De‑risked to answer directly, without the reflexive refusals of over‑aligned models. You decide what it works on.
🧠 Reasons in the open	Native <think>…</think> chain‑of‑thought distilled from a panel of frontier teachers — transparent, inspectable reasoning you can actually read.
💻 Codes and reasons	Tuned for software engineering, step‑by‑step problem solving, and precise technical explanation.
👁 Sees	A full vision encoder — text and images in one 9B model.
⚡ Runs anywhere	BF16, NVFP4 for Blackwell, and a complete GGUF quant ladder with the MTP speculative‑decode head.

📦 This repository — BF16

Full‑precision (bfloat16) weights: the reference model — ideal for fine‑tuning, evaluation, vision‑language use, and as the source for further quantization. Keeps the vision encoder and the multi‑token‑prediction head. Full 1M context.

⚡ Quickstart

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "Blackfrost-AI/PINQWEN-3.5-9B-1M-BF16", dtype="bfloat16", device_map="auto")
proc = AutoProcessor.from_pretrained("Blackfrost-AI/PINQWEN-3.5-9B-1M-BF16")

messages = [{"role": "user", "content": "Write a Python LRU cache with O(1) get/put and explain the design."}]
ids = proc.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=1024)
print(proc.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

🎛 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

🎯 Intended use

General reasoning, coding assistance, technical Q&A, ultra‑long‑document and whole‑codebase understanding, and multimodal (image + text) work. As an uncensored model, it will follow instructions directly — use responsibly and verify outputs for high‑stakes tasks.

📄 License

Apache‑2.0. Base architecture: Qwen 3.5 9B.

---

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