Intern-S2 Official Release Adds 397B Multimodal Weights; Legacy Preview API to Retire at End of October

Shanghai AI Laboratory’s InternLM released the production version of Intern-S2-397B on September 13. This is not simply a renaming of the July Preview: according to the organization, the new release retains its scientific multimodal reasoning capabilities while expanding agent training for complex, long-horizon tasks. Although the model is branded as 397B, an inspection of its Hugging Face files indicates approximately 403B parameters. The public release includes BF16 and FP8 weights under the Apache 2.0 license.

The training design begins by using rendered pages from scientific papers directly as visual data, enabling the model to jointly learn text, formulas, charts, and layout relationships. This avoids breaking those correspondences through preliminary OCR or document parsing. Post-training covers more than 20 scientific domains and connects multiple agent frameworks to sandboxes for black-box agentic RL training on multistep tool use. Text evaluations allow inference with up to 256K tokens, while visual evaluations are capped at 64K. These settings mean leaderboard results should not be treated as representative of short-context, low-cost deployments.

The engineering support is more complete than in the Preview. The model card provides deployment instructions for LMDeploy, vLLM, and SGLang, along with an OpenAI-compatible API, tool calling, and thinking mode. LMDeploy can also expose an Anthropic-compatible endpoint for use with Claude Code. The official `intern-s2` API offers a 256K context window and enables deep thinking by default; the documentation explicitly advises against disabling it for agentic workloads that depend on tools.

Migration risks should be addressed as a priority. `intern-s2-preview-397b` is scheduled to be discontinued on October 31, while the easily confused `intern-s2-preview` actually points to the smaller 35B-A3B version. One section of the documentation says `intern-latest` already points to the production release, but another model table still lists the Preview. Production systems therefore should not rely solely on the alias; they should query the models API for `ref_model` and pin an explicit ID.

Public results currently still come primarily from the development team, and some scientific datasets and long-horizon agent environments lack independent reproduction. The 397B weights are also far beyond the capabilities of a typical workstation. The next points to watch are quantized releases, consistency across inference frameworks, and reproducible ablation and cost data comparing the production release with the Preview.