InternLM’s 397B Science Model Faces a 35B Value Test

• InternLM has released Apache 2.0-licensed BF16 and FP8 weights for Intern-S2-Preview-397B and offers the model through its own API.
• InternLM’s 35B-A3B alternative has the same listed 256K context window and is already described by the laboratory as comparable to its trillion-parameter predecessor on multiple core scientific tasks.
• The 397B deployment guide recommends eight H100 or eight H200 GPUs, while API pricing, active-parameter count, complete benchmark scores and physical-validation outcomes remain undisclosed in the retained materials.

Shanghai AI Laboratory has released Intern-S2-Preview-397B as a multimodal model for scientific reasoning and long-horizon agents. Organizations can download and self-host the weights or use InternLM’s managed API.

The harder decision is not whether the model is accessible. It is whether the 397B release offers enough advantage over InternLM’s own much smaller S2 model to justify its infrastructure. The published material does not provide a direct, reproducible 397B-versus-35B comparison.

The smaller InternLM model is the missing baseline

InternLM says in its project repository that the 397B model expands visual pre-training, reinforcement-learning coverage and interactive agent environments. It says the model learns from raw pages of scientific literature, jointly trains reinforcement-learning tasks from more than 20 scientific domains, and uses sandboxed environments for long-horizon agent reinforcement learning. These describe the development program; they do not establish reliable autonomous research.

The same repository creates a more demanding comparison than the launch headline. It says Intern-S2-Preview-35B performs comparably to the trillion-parameter Intern-S1-Pro on multiple core professional scientific tasks. InternLM’s API documentation lists that smaller model as 35B-A3B and gives both S2 models a 256K context window.

The 397B-versus-403B discrepancy may reflect different accounting conventions, but the available pages do not explain it. More importantly, they do not disclose the 397B model’s active parameters. Total parameter labels therefore cannot answer how much computation each token invokes or whether the larger S2 model is more efficient than the alternatives.

Architecture claims have no disclosed cost baseline

A launch account describes the 397B model as using a non-Transformer architecture called Mobius with a Memory Decoder. It attributes to Shanghai AI Laboratory a separation of reusable knowledge representations from reasoning operations, plus domain-specific memory modules that can be attached without rewriting the entire base model.

The account also attributes an almost fourfold improvement in end-to-end reasoning efficiency on the same tasks to Mobius. It does not identify the baseline model, hardware, workload, latency, throughput or energy measurement behind that figure. InternLM’s archived model card does not provide a technical report for the 397B architecture, so the claimed gain cannot yet be translated into deployment economics.

The same limitation applies to the claim that encoding scientific documents visually produces one-quarter as many tokens as converting them to parsed text. Token count is a relevant input, but without the visual encoder’s compute cost, training workload and quality comparison it is not a measure of total pre-training expense.

The launch account says the 397B model matched Intern-S1-Pro on core tasks including molecular design and material-structure generation. InternLM’s separate claim that the 35B model is already comparable to Intern-S1-Pro on multiple core scientific tasks narrows the significance of that result. Neither source supplies a like-for-like score table showing what the additional scale buys.

The strongest science result stops at computational screening

The most concrete workflow result concerns a protein binder for IL-7Rα. The launch account says researchers used model recommendations to reduce 43 candidate hotspot residues to 30 and focus them on residues 102 through 189 of the B chain. Under what it describes as the same compute budget, the share of candidates passing AlphaFold3 and Rosetta checks rose from 0.47% to 1.56%.

That is a 1.09-percentage-point increase and about 3.3 times the original pass rate. The result still lacks the number of candidates, repeated-run variance and a control describing how recommendations were selected. It measures passage through two computational filters, not binding in a physical experiment; the account explicitly says the workflow does not replace final biophysical validation.

In a second example, the model generated a proposed structure for the five-element oxide Sr₂Ho₁Cu₂Ru₁O₈ from its formula, including a space group, lattice parameters and fractional atomic coordinates. The account again says physical calculations and experiments remain necessary. Both cases support a role in narrowing candidate searches, not a conclusion that the model can complete scientific discovery autonomously.

The launch account says Shanghai AI Laboratory’s “书生·端砚” platform has been deployed in six areas—life science, critical materials, semiconductors, fusion, quantum research, and earth and weather science—and integrates models, agents, computational tools and experiments. It also says the platform has completed dry-lab-to-wet-lab loops and end-to-end validation in protein and materials work. No site count, study design, physical result or independent evaluation is supplied to show how those deployments perform.

Coding-agent rankings are not model-only measurements

InternLM says it evaluated the 397B model with OpenCompass, VLMEvalKit and AgentCompass, allowing up to 256K tokens for text reasoning and 64K for multimodal tests. The archived repository describes that setup but contains no machine-readable 397B score table.

The launch account says the model ranked above Kimi-2.7-Code and DeepSeek-V4-Pro among open models on TerminalBench 2.1 and SWE-Bench Pro, behind GLM-5.2, and level with GLM-5.2 on SWE-Bench Multilingual. It gives no scores, confidence intervals, agent scaffold, resource limits, retry policy or failure treatment. These are therefore launch-account rankings attributed to the laboratory, not an auditable measurement of the model alone.

Two retained benchmark studies show why those missing controls matter. In a Terminal-Bench 2.0 experiment, Anthropic held the Claude model, harness and task set constant across six resource configurations. Moving from strict limits to uncapped resources increased the score by six percentage points and reduced infrastructure errors from 5.8% to 0.5%. The study concerns an earlier Terminal-Bench version, so it does not invalidate InternLM’s result; it shows that hardware enforcement is part of the test.

OpenAI’s audit of SWE-Bench Pro examined the 731-task public split. Its automated analysis identified 200 tasks, or 27.4%, as broken, while its human campaign identified 249, or 34.1%; OpenAI estimated that roughly 30% of the benchmark was broken. The cited problems included overly strict tests, underspecified prompts, low-coverage tests and a misleading prompt. That finding challenges the benchmark’s resolution, not any one model’s result.

Open weights still require concentrated infrastructure

The project is released under Apache 2.0, and InternLM publishes both BF16 weights and a separate FP8 variant. The repository supports LMDeploy, vLLM and SGLang, while the official API offers an OpenAI-compatible endpoint. At the snapshot, the model pages said no Hugging Face inference provider had deployed either artifact, leaving InternLM’s service as the documented managed option.

InternLM’s deployment guide recommends H100 or H200 nodes with eight GPUs. It documents tensor or expert parallelism and long-context configurations, but provides no throughput, latency, utilization or cost result. The retained API material also gives no price. Downloadable weights shift operational control to the user; they do not make a 397B deployment inexpensive.

There is an additional routing hazard. The intern-latest API alias points to intern-s2-preview-397b, while the legacy intern-s2-preview identifier still points to intern-s2-preview-35b. Applications that use the family name rather than an explicit version can therefore invoke a different model than intended.

The launch account says the 397B model is being jointly optimized with Huawei’s Ascend ecosystem. The available material provides no Ascend configuration, benchmark or cost comparison. The concrete deployment recommendation is currently framed around Nvidia H100 and H200 hardware, leaving the practical extent of Ascend support unresolved.

The next test is 397B against 35B under one harness

The decisive evidence is a controlled comparison between the two S2 models. InternLM would need to publish their exact scientific and agent scores under the same prompts, scaffolds, inference budgets, hardware limits and retry rules, along with the 397B model’s active-parameter count. Hardware results should report latency, throughput, memory, utilization and cost for both BF16 and FP8 deployments, while the hosted API needs a price and rate limits.

Scientific validation needs a different denominator: candidate counts, independent repeats, failure distributions and physical or wet-lab outcomes against a stated baseline. Deployment claims for the six research areas also need measured outcomes rather than a list of fields.

Until those comparisons exist, Intern-S2-Preview-397B is a consequential distribution release and a testable scientific-agent preview. The evidence does not yet show that its additional scale produces enough capability or efficiency to beat InternLM’s smaller substitute in a real research budget.