Formal Proofs, Arabic Speech, and a Robot With a Map

Not every week is about trillion parameter flagships. This week belonged to the specialists, and honestly these might matter more for what I build day to day.

Leanstral 1.5: Math With Receipts

Mistral released Leanstral 1.5 on July 2, continuing their formal verification line. These models produce machine-checkable proofs rather than confident-sounding prose. Regular LLMs hallucinate mathematics beautifully; formal verification models show their work in a language a proof checker can validate.

My use case is humbler than theorem proving: verifying that the locking logic in my game backend can't deadlock. I described the lock ordering rules, and the model walked through the invariant checks formally instead of vibing its way through an explanation. Did I trust it blindly? No. Did it find an ordering ambiguity I'd glossed over? Absolutely yes.

Sometimes you don't need a smarter answer. You need an answer that comes with a proof.

Robostral Navigate: 8 Billion Parameters and a Camera

Also Mistral, also this week: Robostral Navigate, an 8B vision model for single-camera robot navigation. Feed it one camera stream, it reasons about where it is and where to go.

• Single camera, no lidar, no fancy sensor arrays
• Small enough that onboard compute becomes realistic
• Part of Mistral's quiet expansion into embodied AI

Do I have a robot? No. Do I have a shopping list and a dream involving a warehouse inventory rover for the homelab? Obviously. Watched every demo clip twice.

Cohere Transcribe Arabic

Cohere released Transcribe Arabic on July 7: a 2B parameter speech-to-text model under Apache 2.0 that outperforms Whisper v3 Large across Arabic dialects. It continues a pattern worth noticing: Transcribe in March, Command A+ in May, North Mini Code in June. Three permissively licensed releases in six months, with more promised.

My test was selfish and immediate. I record my classes to review later, so I ran a lecture recording through it alongside my usual Whisper setup. On English it held its own. On the bits where my teacher slips into Arabic to explain a concept properly, Whisper produced confident nonsense while this thing just... transcribed it.

The models that win aren't always bigger. Sometimes they're just aimed at people the big labs forgot.

Why Specialists Matter

• They run on hardware you already own
• They solve one problem completely instead of ten partially
• Permissive licenses mean no invoice anxiety
• They fill gaps the frontier labs consider too small to bother with

Tomorrow, the biggest model of the year goes general availability, and the cycle starts again. But keep an eye on the weird ones. That's where the unexpected careers get built.

Cheers, Yassen