Scientific workflows often fail not because a model cannot interpret an input, but because it cannot connect that interpretation to the next experiment or verify whether an action worked. This model’s central bet is to bridge scientific perception and agent execution: scientific evidence, software feedback, and natural-language instructions all inform the same iterative process.
Key Capabilities
- Scientific inputs beyond images and text: Encoders and projectors bring data such as molecular graphs, crystal structures, biological sequences, and spectra into a shared language-model context. This makes predictions, site localization, and structure inference possible without reducing every input to a text description first.
- Analysis that can be checked: The model can use specialized software, inspect intermediate outputs and visual states, then adjust its next action. That loop is useful when a research result depends on a sequence of tool-mediated steps rather than a single answer.
- A general agent backbone: Its 320-billion-parameter mixture-of-experts backbone is trained for tool use, search, coding, and scientific workflows. Reported evaluations show strengths in machine-learning engineering and professional document tasks, alongside competitive results in computer use and research execution.
Who It’s For
A good fit for researchers exploring agents that work across scientific data and specialized software, especially when intermediate results can be verified. Look elsewhere if you need a lightweight model: the full release is roughly 645 GB in bfloat16 and its deployment examples target multiple high-memory GPUs. Scientific inputs also require the relevant encoders, projectors, and preprocessing—not just the base model.