Macaron-V1: Open Continual Learning Agents with Mixture-of-LoRA
mindlab-research · hf · 2026-08-11
Macaron-V1 is an open agent-model family focused on experiential intelligence, learning from real environments and continuing to improve post-deployment.
- Architecture: Uses a Mixture-of-LoRA (MoL) architecture that freezes a base model, composes specialist LoRA adapters, and selects one per turn. The flagship Venti combines a 744B GLM-5.2 base with four LoRAs (chat, agent, coding, GenUI); the Tall version uses a 50B Qwen3.6 for local deployment.
- Mechanisms: Achieves adaptation and collaboration through Model-Harness Co-design and recursive self-improvement loops. Includes the UI4A GenUI harness, stateful action substrate, and the MindForge agentic RL framework.
- Infrastructure: Supported by the MinT post-training platform, LongStraw long-context RL method, and stability techniques for sparse MoE.
- Evaluation: Validates the current system across benchmarks, though compounding gains from continual learning remain an open question.
More from Models
- Local Open-Source Model Self-Checks and Runs Code: What Do Subscriptions Still Buy? — truecakesnake · 2026-08-11
- OpenAI Accused of Stifling Model Creativity via Strict Temperature Controls — RileyRalmuto · 2026-08-11
- GLM-5.2 Pricing Drops: High-Level Intelligence Gets Dramatically Cheaper — tobowers · 2026-08-11
- Developer Praises Kimi K3: 'Like if Llama 4 Behemoth Was Real' — willccbb · 2026-08-11
- Alibaba Confirms Qwen3.8-27B Open Weights Landing This Week — max_paperclips · 2026-08-11
- Abacus AI Releases Smaug-Agentic, Topping Open-Source Leaderboard for Agentic Coding — bindureddy · 2026-08-11