Macaron-V1 review: a 748B personal-agent model with 2M context and modular LoRA experts
AI寒武纪 · wechat · 2026-07-22
This long review tests Macaron-V1, a “personal agent” model from MindLab, and argues that it is meaningfully different from the usual chat-with-plugins pattern.
What it is
- Macaron-V1 comes in two versions:
- Venti: 748B parameters, built on GLM-5.2 post-training, with native 2M context.
- Tall: 35B, based on Qwen 3.6 post-training, aimed at local deployment and privacy-sensitive use cases.
- The architecture uses MoL (Mixture of LoRA): a frozen GLM-5.2 base plus four 1B LoRA experts for Chat, Agent, Coding, and UI/A2UI.
What the reviewer tested
- UI4A / Artifacts: Macaron can render interactive dashboards rather than static answers. The reviewer asked for a personal finance console with editable tables, linked charts, and a humorous assistant avatar; the generated UI updated live when data changed.
- Coding: In Claude Code, the model was used to build a self-contained Three.js orrery and an interactive 12-part ink-wash scroll. The author says Macaron repeatedly ran browser-based checks and fixed rendering issues during construction.
Main takeaway
The article’s core claim is that Macaron-V1 is not just a chat model with tools, but a modular agent system that can span conversation, coding, UI generation, and local deployment in one stack.
More from coding & agent
- Two papers use LLMs to improve retrieval indexing and grounded answers — _reachsumit · 2026-07-22
- Grok Build turns one prompt into a full ARPG with AI-generated game assets — tetsuoai · 2026-07-22
- A dad built a controller-ready game in two hours with Grok 4.5 — minchoi · 2026-07-22
- Atomic-Chat pitches a fully offline open-source ChatGPT alternative — rohanpaul_ai · 2026-07-22
- ComfyUI Wan dance test renders a 30-second clip in 70 minutes — tostane · 2026-07-22
- Claude Code adds iOS Simulator control for side-by-side mobile testing — xiaohu · 2026-07-22