Xiaohongshu (RedNote)’s dots infrastructure team has published BigMac as an open-source multimodal training system designed to remove a common tradeoff: train faster at the cost of activations overflowing GPU memory, or conserve memory and leave expensive chips idle. BigMac preserves the LLM pipeline while arranging vision encoders and media generators around it to reduce idle time and redundant activation storage.
How it works and measured results
According to the team’s reported benchmarks on Qwen3-30B-A3B workloads, BigMac increased training throughput by between 1.08× and 1.9× while keeping peak memory usage stable as batch sizes grew. Rather than relying on larger activation buffers, the system addresses the inefficiency through scheduling and component nesting around the core LLM.
Practical implications
Multimodal AI development has often led to infrastructure waste: labs bought extra GPUs to compensate for pipelines that caused chips to wait, and those additional costs were ultimately reflected in prices charged to customers (for example, token prices). By turning idle GPU time and bloated activation storage into a scheduling problem, BigMac could allow the same cluster to train larger multimodal models faster, run more workloads concurrently, and reduce the infrastructure cost per generated token if integrated into commercial stacks.
Key takeaway
BigMac itself does not create cheaper intelligence by inventing new algorithms; it identifies and reduces GPU inefficiencies that have been inflating the cost of multimodal training. If the approach proves robust in production environments, it could materially lower infrastructure costs behind model training and token generation.
