AINEWS
Mind Lab claims 6 of 12 benchmark wins against GPT 5.4 by training 4 billion parameters on top of a frozen 748 billion base, and Meituan just led a $50M round to back the bet.
A thirty-person team in Beijing says it matched or beat GPT-5.4 on six public benchmarks without retraining the base model. The flagship, called Macaron-V1 Venti, is 748 billion parameters total, but only four billion of those are Mind Lab's own work. The rest is a frozen copy of GLM-5.2, the open-weight base model from Zhipu AI.
What the other four billion does is the story. Mind Lab trains a set of small, cheap patches called LoRAs, or low-rank adapters, the same technique used to fine-tune Stable Diffusion and open Llama variants. The team trains one adapter per task: chat, agentic tool use, coding, and UI generation. At inference time, the system swaps in the right adapter for the task and accumulates per-user adapters over time. The company calls this mixture-of-LoRAs, or MoL. It is, in effect, a way to keep the base model still and let the skill layer move.
In a 36kr founder interview and a Mind Lab research blog post, Mind Lab claims six state-of-the-art results across twelve published benchmarks for the official Venti release, which shipped July 21. The other six are described as close to frontier. The weights are published on Hugging Face, but the technical report is not. That means the benchmark tables are company-asserted until an outside lab reproduces them.
For a smaller, local-deployable option, Mind Lab released Tall the same day. It is 50 billion parameters built the same way on top of Qwen3.6, Alibaba's open base, and is aimed at running on a single high-end workstation. Both versions claim a native two-million-token context window.
Co-founder Andrew Chen co-authored the 2023 FireAct paper with Shunyu.
Mind Lab closed a roughly $50 million A round in early 2026, according to Sina Finance, Geekpark, and Eastmoney, all of which trace back to the same company announcement. Meituan Strategic Investment led, with Yuanhe Puhua, headphone maker Shokz, and Variable Capital participating. Total raised, including earlier rounds, is about $60 million; historical shareholders include Ant Group, Sequoia China, ZhenFund, and Gaorong. Mind Lab also says Macaron-V1 hit $10 million in annualized recurring revenue inside roughly two weeks of commercialization, another company-asserted number, not yet audited.
The infrastructure is the part that may outlast the benchmarks. In December 2025, the team says it ran a trillion-parameter reinforcement-learning job using a custom mixed-coordination parallel engine plus a truncated importance-sampling step designed to keep the training and inference distributions aligned. In January, it released MinT, a platform for serving and managing LoRA adapters at scale, claiming to handle more than a million adapters and a roughly tenfold speedup in loading them in real time. Both numbers come from the same founder interview and have not been independently reproduced.
If a thirty-person team can move six benchmarks by changing 0.5% of a frozen 748-billion-parameter giant, the binding constraint for the next capability jump is no longer who can afford the next pretraining cluster. It is post-training method, adapter infrastructure, and who controls the serving stack. The falsifier is simple: if OpenAI, Anthropic, or Google were already running this trick at scale, it would be visible in their public model cards. Right now it is not, which is why a Meituan-led check makes sense as a hedge on the post-training layer.
The open question is reproducibility. The weights are out, the report is not, and the benchmark tables come from the team that shipped the product. Until an outside lab runs the same evaluation suite and gets the same numbers, the Venti results should be read as a strong company claim, not a settled leaderboard position. Mind Lab says the technical report is coming; for now, the bet and the check are running ahead of the proof.