Mistral-Medium-3.5-128B is a large language model released by Mistral AI, released under the Modified MIT license, with a claimed knowledge cutoff of 2026-03-31. Its weights, configuration, and tokenizer files have been cryptographically hashed and are queued for on-chain attestation by Weight Registry.
Description
128B dense model merging what were previously three separate specialist lines (Mistral Medium 3.1 for chat, Magistral for reasoning, Devstral 2 for coding) into one set of weights -- notable both as a rare open-weighting of a 'Medium'-tier flagship and for the merge-not-just-scale design choice. 256k context, accepts text and images.
Metadata
Hugging Face stats
111,223 downloads on Hugging Face · 459 likes
Weight Hashes
Weights file
SHA-2563e9870348e3b7d3cfa1c248dd38b98cae6b0083d03de44d4b7f77f7f14e8b911
SHA3-256f1434e8e433db0863138676b5ea7da904239172519ec38e25d5d999e8beb08cb
BLAKE35b50c6b6fe22e9b3907ece08c2b345fd42e0adbe31b0fdfb253190d85d419690
Config file
SHA-2566ebadc48e3bdfff1a6a0b863530c68c9725a49eecd21c0527bdcb21b822f4d71
SHA3-2567b1df84778edbc55d7b66b715f7003775f57ec0c1163f4ae916521c54485de0d
BLAKE3297ab1873cce3bfdfe8214ced265cd0ef6971505220d04a99ff32f52ddf2ab22
Tokenizer file
SHA-2562ba5b3330fd84d5376fcca797cfb3b42eee6241ce23e3271e6fb2a115a8751bd
SHA3-2565d417423c177ba0bfc8ed931c5c4f381ea653c350f27381d5cda9f2370aefb61
BLAKE3aeb924c043d10ca77948506e3a32a9098d390db791446b81e0554e36c9a89c7f
Verify it yourself
Download the file(s) from the source below, then hash them locally (sha256sum <file> on Linux/macOS, certutil -hashfile <file> SHA256 on Windows) and compare the output against the hash(es) shown above.
Anomalies Flagged
- MULTIPLE SHARD SETS found (likely different precisions/formats): chose prefix 'model' (3 shards), did NOT hash other shard set(s): {'consolidated': ['consolidated-00001-of-00003.safetensors', 'consolidated-00002-of-00003.safetensors', 'consolidated-00003-of-00003.safetensors']} -- VERIFY the chosen precision is the intended default
- other weight-like files present but not hashed: ['consolidated-00001-of-00003.safetensors', 'consolidated-00002-of-00003.safetensors', 'consolidated-00003-of-00003.safetensors']