My biggest issue with Devstral and even their biggest model is that they’re dangerous unless closely directed and reviewed and i mean CLOSELY. Unfortunately mistral models will believe and do anything.
I don't really see the novelty, you can achieve the same on any VM with SQLite + Websockets for realtime sync of state - and if start time is an issue or even deploy, Alpine linux is just 5MB..
The actual interesting part is the scaling, from 0 to 100 instantly.
depends on usage, Gemma 4 is better on visuals/html/css and language understanding (Which probably plays a role in prompting). But it's worse at code in general compared to Qwen 3.5 27B.
Great writeup, I’ve been running my own mitm measuring token usage on antigravity since release. In the past week and with the introduction of ”credits”, the amount of usable opus on the 350$ ultra plan was >4000$ it is now <1200$ which means around 4x less usable tokens.