More posts
Ran the numbers against DeepSeek-v4-Pro. Luna wins for what I need, so I'm moving my subagent slot over to it. Setup right now: Grok-4.5 or Kimi-K3 orchestrating, Luna doing the actual work.
wrote six years ago into an agent-driven setup.These kinds of projects are a natural fit for agents. Just wrap the core stuff (image find, color find, mouse/keyboard actions) as tools the agent can call ends up way faster than having the agent do real-time vision and control everything itself.
. 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.
