Tryanks Weights Mirror

Model files

This host serves checkpoint shards, tokenizer files, and manifest files over HTTPS. The model list below follows the current Hugging Face models ranking checked on 2026-03-24, with the Qwen3.5 series pinned to the top.

Recommended models
Rank Model Note
1 Qwen/Qwen3.5-35B-A3B Kept at the top. Large Qwen3.5 model for general use.
2 Qwen/Qwen3.5-9B Smaller Qwen3.5 model with lower cost and faster startup.
3 moonshotai/Kimi-K2.5 Large multimodal model with strong current visibility on Hugging Face.
4 MiniMaxAI/MiniMax-M2.5 Large general model with high current traffic on the Hub.
5 mistralai/Mistral-Small-4-119B-2603 Current Mistral entry near the top of the Hugging Face models list.
6 nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 NVIDIA entry that is also near the top of the current list.
Files
Path Size Updated
/weights/qwen3-235b/model-00001-of-00018.safetensors 4.8 GB 2026-03-24 07:42 UTC
/weights/qwen3-235b/model-00002-of-00018.safetensors 4.8 GB 2026-03-24 07:42 UTC
/models/deepseek-vl2/resolve/main/model.safetensors.index.json 19 KB 2026-03-24 07:42 UTC
/assets/tokenizers/llama-bpe-v3.tar.zst 42 MB 2026-03-24 07:42 UTC
/mirror-index.json 214 B 2026-03-24 07:42 UTC
Paths
/models/<repo>/resolve/main/

Use this path for manifest files and other small metadata files.

/weights/<family>/

Use this path for checkpoint shards and merged weight files.

/assets/

Use this path for tokenizer bundles, vocabulary files, and support assets.