MindLab launches Macaron-V1, a 748B open-weight agent model with 2M-token RL
青稞AI · wechat · 2026-07-24
MindLab releases Macaron-V1, an open-weight agent model family
Macaron-V1 comes in two versions: Venti, a 748B open-weight flagship built on a GLM-5.2 base with four 1B LoRA experts, and Tall, a lighter local-deployable version built on Qwen3.6 with four 3.7B LoRA experts.
What’s new
- Uses a Mixture-of-LoRA (MoL) architecture so specialized capabilities can be trained, updated, and plugged in independently.
- Introduces LongStraw, which extends reinforcement learning training to 2 million tokens of context.
- Trains the model around tool use, terminal operations, software engineering, and UI generation, with a harness designed to match production deployment.
Claimed results
- MindLab says Macaron-V1 can match or exceed Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on many of its benchmarks.
- It highlights especially strong performance in life-like tasks, coding, and generative UI.
Release details
- All versions are open-weight on Hugging Face.
- An API is available for Venti, plus a coding-focused version and a local-deployable Tall variant.
- The team also open-sourced its long-context RL paper and a plugin for rendering UI4A outputs.
Related event: MindLab Releases Macaron-V1 Open-Weight Agent Models(3 posts)→
More from coding & agent
- ChatGPT Voice in Codex behaves like an orchestrator thread, but needs explicit thread switching — petergyang · 2026-07-24
- Kimi K3 Max matches GPT 5.6 Sol Max on software tasks at 55% of the price — togethercompute · 2026-07-24
- Official repo for Designing Multi-Agent Systems ships PicoAgents and full implementations — tom_doerr · 2026-07-24
- Code and Codex hide a surprising amount of complexity for new users — emollick · 2026-07-24
- New ComfyUI tile-refine node promises faster, cleaner image upscaling — blakeem · 2026-07-24
- Don't Let Models Write Their Own Audit Logs in Agent Stacks — anp2_protocol · 2026-07-24