InternLM has introduced Intern-S2-397B, a 397-billion-parameter multimodal foundation model aimed at scientific intelligence and long-horizon agents. The distinctive point is not just scale, but a change in training paradigm: the model learns directly from raw pages of scientific literature, without intermediate parsing. Text and visual relationships are modeled in the same representation space, preserving the correspondence between symbols and figures and strengthening spatial reasoning. For those building scientific pipelines, this approach reduces dependence on layout extractors and imperfect transcriptions, but shifts complexity to corpus quality and the infrastructure required for this kind of training.

On the reinforcement learning side, InternLM has extended coverage to more than twenty scientific domains, jointly training tasks such as biomolecular interaction design and material structure generation. The team's note describes competitive general-reasoning performance among open-source models and strong results on specialized tasks, without providing specific benchmarks. The choice of domains is nonetheless telling: the goal is not a scientific chatbot, but an engine capable of operating on problems that require hybrid representations, where the visual component is not an accessory.

The agent side follows the same logic. Intern-S2-397B connects multiple agent frameworks to large-scale sandboxed environments for black-box reinforcement learning. The aim is to improve generalization and raise the capability ceiling for long-horizon tasks. Structurally, this signals a shift from static evaluations to interactions in controlled environments: for organizations with proprietary scientific data, the ability to train and validate agents in local sandboxes becomes a control and security element, not a marginal detail.

For those evaluating on-premise deployment, a 397-billion-parameter model is not a neutral object. It requires thinking about aggregate VRAM, inference throughput, and quantization strategies. A single consumer card is not enough: a cluster or hybrid infrastructure must be designed to serve multimodal and agentic workloads with acceptable latency. The absence of hardware details in the announcement leaves the question open, but the signal is clear: new-generation scientific models push toward on-premise configurations for data sovereignty and control reasons, with non-trivial capital costs. AI-RADAR explores these trade-offs in the /llm-onpremise section, without recommending a single choice.

The beneficiaries of this approach are research groups and companies with rich scientific archives that are difficult to move to the cloud. The potential losers are vertical tools that currently sell extraction and annotation pipelines for scientific documents, because part of the parsing logic is absorbed by the model. For hardware vendors, demand shifts toward systems with high video memory and fast interconnects, needed to host models of this size without sacrificing agent latency. InternLM's bet is that removing intermediate parsing and jointly training on heterogeneous domains produces more robust generalization than stitching together specialized components.