Views
No views yet

🪟 A 9B that reads a million tokens, thinks before it speaks, and never refuses — run it on your own machine.
<think> → answer), across a full one‑million‑token context window. This is the complete GGUF drop for llama.cpp — every quant, the vision projector, and the MTP speculative‑decode head for faster generation.<think>…</think> chain‑of‑thought from a panel of frontier teachers.mmproj-*) for image input in compatible runtimes.| File | Precision | Size | Notes |
|---|---|---|---|
PINQWEN-3.5-9B-Q2_K.gguf | 2‑bit | ~3.7 GB | Smallest; tight budgets |
PINQWEN-3.5-9B-Q3_K_M.gguf | 3‑bit | ~4.5 GB | |
PINQWEN-3.5-9B-Q4_K_M.gguf | 4‑bit | ~5.4 GB | ⭐ Recommended balance |
PINQWEN-3.5-9B-Q5_K_M.gguf | 5‑bit | ~6.2 GB | Higher quality |
PINQWEN-3.5-9B-Q6_K.gguf | 6‑bit | ~7.1 GB | |
PINQWEN-3.5-9B-Q8_0.gguf | 8‑bit | ~9.2 GB | Near‑lossless |
PINQWEN-3.5-9B-f16.gguf | f16 | ~18 GB | Full precision |
mmproj-PINQWEN-3.5-9B-BF16.gguf | — | ~0.9 GB | Vision projector (image input) |
1# text
2llama-cli -m PINQWEN-3.5-9B-Q4_K_M.gguf -p "Explain the CAP theorem." -ngl 99
3
4# thinking off (direct answers)
5llama-server -m PINQWEN-3.5-9B-Q4_K_M.gguf -ngl 99 \
6 --chat-template-kwargs '{"enable_thinking":false}'
7
8# with vision
9llama-mtmd-cli -m PINQWEN-3.5-9B-Q4_K_M.gguf \
10 --mmproj mmproj-PINQWEN-3.5-9B-BF16.gguf --image pic.jpg -p "Describe this."| 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 |