Venti/Coding-Venti: 748B aggregate (744B base + four 1B LoRA specialists), ~40B active; Tall: 50B aggregate (35B-A3B base + four 3.7B LoRA specialists)
1M tokens (Venti and Coding-Venti); 262K tokens (Tall)
MIT
Official non-preview Macaron-V1 release grouping three open-weight text-generation checkpoints: the flagship Macaron-V1-Venti, the directly usable coding-specialized Macaron-V1-Coding-Venti, and the smaller Macaron-V1-Tall. Venti is described as the first post-trained derivative of GLM-5.2 and combines its sparse Mixture-of-Experts base with four 1B-parameter LoRA specialists, while Tall applies the same system to Qwen3.6-35B-A3B using four rank-64 LoRA specialists of approximately 3.7B parameters each. The Mixture-of-LoRA architecture divides work among L0 Chat, L1 Agent, L2 Coding, and L3 GenUI specialists; L0 routes each new request, ongoing reasoning and tool interactions remain with the selected specialist, and concise summaries communicate completed work between their separate contexts. Coding-Venti merges the coding specialist into the GLM-5.2 base so it can be loaded as a conventional coding checkpoint without the routed multi-LoRA harness. Venti and Coding-Venti support up to a 1M-token context, while Tall's released configuration supports 262K tokens. Post-training used Mind Lab's MinT infrastructure and MindForge recursive self-improvement loop, with LongStraw providing an experimental execution path for reinforcement learning beyond two million tokens under fixed GPU budgets. Mind Lab evaluated V1 across ChatBench, LivingBench, VitaBench, PinchBench, SWE-bench Verified, DeepSWE, SWE Atlas QnA, TerminalBench 2.1, and UI4A-Bench, claiming that Venti matches or exceeds the listed proprietary frontier baselines across most tracks while remaining closest to, rather than ahead of, the frontier on coding. The release includes BF16 base weights and adapters, chat templates, an MIT-licensed OpenAI-compatible routing proxy, and deployment profiles for native multi-LoRA serving through vLLM or SGLang.