MacaronV1 review says UI, agents, and 2.1M-token context are the real story
卡尔的AI沃茨 · wechat · 2026-07-24
What the article covers
A long hands-on review of MacaronV1 (built on GLM-5.2) argues that its most interesting trait is not benchmark hype, but stability in real workflows: UI generation, agent-style chat, coding, and long-context use.
Key points
- The author first compares MacaronV1 against other models on UI-centric tasks using Claude Code + a TasteSkill-like setup, and says the model can generate a polished, production-looking web demo from a short prompt in about 10 minutes.
- A separate multi-agent chat room demo also works well, with different agents taking on roles like coordinator, researcher, educator, and engineer.
- On a physics page task, the model builds an interactive gravity particle lab in plain HTML/CSS/JS with stable particle motion, dragable gravity source, repulsion mode, and bounded velocities.
- On a classic 2048 task, it reportedly gets the merge rules, invalid-move handling, score tracking, win/lose states, and touch/keyboard input right in one pass.
- The author then describes three technical ideas from the Macaron team:
- MoL: freeze the base model and do all post-training in LoRA space, so different task families can specialize without overwriting one another.
- A2UI: a plugin for evaluating whether a model can generate useful UI components inside a conversation.
- LongStraw: a training method for extending context by reading the prompt once and replaying only the reasoning trajectory, reducing memory pressure.
- The article claims 8× H20 can train a 27B model to 2.1M tokens, and 32× H20 can run a full GLM-5.2 78-layer / 256-expert MoE model at the same length.
- It also mentions MindForge, which mirrors the real agent environment during training, and a recursive self-improvement loop where the model generates tasks, solves them, audits traces, adjusts the harness, and trains on the improved data.
Bottom line
The post’s main takeaway is that MacaronV1 is interesting less as a raw benchmark winner and more as a model that appears unusually strong at UI, agentic interaction, and long-context engineering—with an emphasis on deployable, stable outputs rather than leaderboard theatrics.
More from Models
- Claude screenshot says it has no original math results, and the reply is pure meme fuel — mimi10v3 · 2026-07-24
- Google’s internal code still outclasses Gemini on writing quality, says engineer — yunta_tsai · 2026-07-24
- Anthropic is reportedly hiding Fable 5’s reasoning traces — ns123abc · 2026-07-24
- MindLab open-sources Macaron-V1, a GLM-5.2 family built on continual learning — 机器之心 · 2026-07-24
- Unlimited-OCR returns to No. 1 on Hugging Face trending models — _akhaliq · 2026-07-24
- V4 update slips, with K3 weights possibly landing in the same release window — teortaxesTex · 2026-07-24