Atlas

Models

← All models

Macaron-V1-Venti

Mind Lab · 2026-07-21 · 784.1B parameters

Macaron-V1-Venti is Mind Lab's flagship open-weight personal-agent model, built on a frozen GLM-5.2 base with four LoRA specialists for chat, agent tasks, coding, and code-native Generative UI. An L0 router selects a specialist for each request, and the model was trained with its production-aligned agent harness. Mind Lab markets it as 748B (a 744B base plus four 1B specialists), while the published safetensors contain 784,082,110,464 parameters across the base checkpoint and adapters.

Benchmark scores

BenchmarkScore
ClawGym-Bench77.7
DeepSWE v1 (Best of up to 3 attempts)58.4
Macaron ChatBench58.3
Macaron LivingBench64.0
PinchBench Best Score94.0
SWE Atlas - Codebase QnA (pass@3)49.5
SWE-bench Verified85.6
Terminal-Bench 2.187.6
UI4A-Bench87.8
VitaBench60.0
VitaBench 2.0 - Rewrite (Agentic Memory)46.0
τ³-bench average69.3
Loading Atlas data…