More posts
I posted in July that once agents get everywhere, subscription skills would probably show up. Felt speculative at the time. Then Salesforce did it in September: they put 37 sales skills inside Claude, and every time a skill successfully calls Salesforce, that's one interaction. Salesforce says they're not charging for those interactions right now. Free today. Paying later, apparently.
Most desktop apps in this space are dropping the coding part and turning into general office tools. Command Code's desktop is more of a coding IDE. Classic chat pane, plus live agent activity so you can see which files it read and which commands it ran. Built-in browser and Design mode: click an element on the page and hand it to the agent to change. It ships with taste, a coding style that keeps learning your habits, and it picks up skills and MCPs globally on its own. Right now taste is the only thing here that stands out, and memory can do the same job just as well. 🔗 https://commandcode.ai/desktop
. 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.
