More posts
stress tested DeepSeek-v4-flash-0731 full-stack coding — built a macOS Dynamic Island app. media controls, file stash, system info, all docked in the notch. backend is on par with GLM-5.2. frontend? zero chance. no vision modality means it's coding blind. setup: Pi native coding agent, no skills, no subagents, no extensions. one model, one session, whole project. features took 2 hours — 1M+ tokens, 99% cache. UI polish (rounded corners, animations) took a full day. ~40 RMB total. do NOT use DS for UI animation work. one animation stumped me for half a day. gave up, opened Grok-4.5, 10 min done lol open sourced, grab and build on it: https://github.com/haoyiyin/dao
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.
. 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.
