What it is
A 397B scientific multimodal agent model trained with unified SFT+RL+agentic RL, plus time-series forecasting and a plug-in memory path for domain specialization. Unlike static scientific QA models, it targets long-horizon tool-grounded science work: multimodal scientific pretraining plus unified reasoning/agentic RL, with a forecasting branch for time series and plug-in memories that specialize a frozen 397B backbone.
Key results
- Biology-Instructions avg improves 56.92→60.32 with Intern-MemDec-4B
- Time-series encoder supports 300k steps vs ~240k previously
- At max length, inference is ~5-6× faster and uses ~20% of prior GPU memory
Scientific foundation models can add domain expertise and numerical forecasting without rewriting the main backbone by combining agentic post-training with modular memory extensions.