Mistral Medium Rumored at 128B

Hot take: this is less about the raw parameter count than Mistral trying to reclaim the “practical frontier model” slot with something that feels big enough for serious work but still deployable.

• –Mistral Small 4 is officially 119B parameters with 6.5B active, so a 128B Medium would probably be another MoE-style design, not a straightforward dense jump.
• –The Reddit thread cites a vLLM PR, which makes this feel like a tooling breadcrumb rather than an official launch announcement.
• –If Mistral ships a Medium successor that is meaningfully better on coding and reasoning, that would strengthen its position against bigger but pricier closed models.
• –For local and self-hosted users, the key question is active parameters and memory footprint, not total parameter count.
• –The naming cadence suggests Mistral is continuing a rapid family refresh across Small, Medium, and the surrounding tooling ecosystem.