More posts
. just needed a multimodal model that's actually good AND doesn't bleed me dry. tried everything on the shelf, only Xiaomi's MiMo-V2.5 fit lol every popular LLM rn is basically a one-trick pony — killer at one thing, mediocre at the rest. you really can't have it all.
. Every account gets a cloud box. You can build projects on it, phone to web and back. Tried it asked Grok to make a pixel baseball game right in the app. It showed exactly where the code hit disk. Sent me test links sometimes, ran the HTML in the chat window other times. Either way, game ran. Wild compared to just running skills on their server before.
. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.
It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
