Model Release

InternLM Sage 2.1 Tops MLE-Bench at 90.3 With MIT License

Shanghai AI Lab's 320B open-weight scientific model scores 90.3 on MLE-bench Lite, handling proteins, crystals, and molecular graphs under MIT.

Share:XLinkedIn

Key Takeaways

  • 320 billion total parameters with 18 billion active per token in a mixture-of-experts architecture under MIT license, making Sage S2.1 Flash the largest open-weight scientific AI model published to date
  • 90.3 on MLE-bench Lite, ahead of DeepSeek-V4.1-Flash-1017 at 89.2, with first place on 4 of 15 benchmarks and second on 7, per InternLM's own evaluations with no independent reproductions yet
  • 10-plus scientific input types including protein sequences, molecular graphs, 3D protein structures, crystal structures, and DNA sequences, unified through the Sage SciPerception architecture alongside natural language
  • 1,048,576-token context window and 645 GB of BF16 weights across 120 shards, with an FP8 variant available, requiring a multi-GPU node with H100 or H200 class hardware to run
  • No safety documentation and no independent benchmark reproductions at launch, with architecture traces pointing to Z.ai's GLM-5.3-Flash base model from August 2026

Shanghai AI Laboratory uploaded the largest open-weight scientific AI model on record on November 15, 2026, with no blog post, no press release, and no mention anywhere on its own homepage. The Hugging Face model card contained a single line when it first appeared: the MIT license. By the time a GitHub Pages microsite explaining the release appeared the following day, 645 gigabytes of model weights had already been downloading for 24 hours. The name is Sage S2.1 Flash. The claim is a score of 90.3 on MLE-bench Lite, the highest any model, open or closed, has reported on the benchmark that measures AI agents on real machine learning engineering tasks pulled from live Kaggle competitions.

Sage S2.1 Flash Scores 90.3 on MLE-Bench Lite, Outpacing DeepSeek by 1.1 Points

MLE-bench Lite, first published by OpenAI in 2024 and presented at ICLR 2025, is built from 75 Kaggle machine learning competitions. It tests whether an AI agent can build, train, and iterate on a model well enough to earn a Kaggle medal, which requires outperforming a majority of human competitors on real data science problems with real datasets. When o1-preview first ran MLE-bench at launch, it hit a bronze-medal rate of 16.9% of competitions. Sage S2.1 Flash's self-reported score of 90.3 on the Lite variant is in a categorically different range, and it is the first open-weight model to surpass 90 on the benchmark by any published report.

According to InternLM's model card and the Sage S2.1 Flash technical report, the model placed first on 4 of 15 benchmarks and second on 7 of 15 across a comparison that included DeepSeek-V4.1-Flash-1017, Qwen3.8-Max, Z.ai's GLM-5.3-Flash, and several GPT and Claude models. The closest competitor on MLE-bench Lite is DeepSeek-V4.1-Flash-1017 at 89.2, a 1.1-point gap. On OfficeQA Pro, Sage S2.1 Flash scores 72.2 for first place. On SpreadsheetBench 2, it scores 37.1. On Terminal-Bench 2.1, it scores 87.9. On DeepSWE, it scores 67.9. These results come from InternLM's own evaluation suite, with no harness files or run logs published alongside them, and no independent organization has reproduced the figures yet.

The model runs on a mixture-of-experts architecture with 320 billion total parameters and 18 billion active per token. That design means inference costs are closer to an 18B model than a 320B one, since most of the weights are dormant on any given forward pass. The full memory footprint is a different matter: 645 GB across 120 shards in BF16 format, with an FP8 variant that cuts that by roughly half. The context window reaches 1,048,576 tokens, just over one million, which allows the model to hold an entire scientific paper, its citation list, supplementary data tables, and a dataset sample all in a single context pass without truncation.

Stay Ahead

Get daily AI signals before the market moves.

Join founders, investors, and operators reading TechFastForward.

A 320B Model That Reads Protein Sequences and Molecular Graphs as Native Inputs

The defining technical feature of Sage S2.1 Flash is a component called Sage SciPerception, a unified input layer that maps more than ten scientific data types into the same embedding space as text. The practical consequence is that the model does not receive a protein sequence or a molecular graph as a text string to be tokenized character by character. It processes them through specialized scientific encoders and trainable projectors, then reasons over them alongside natural language the same way a vision model reasons over image patches as continuous embeddings rather than character sequences.

The supported input types listed in the model card include molecular graphs, crystal structures, protein sequences, 3D protein structures, DNA sequences, RNA sequences, mass spectrometry outputs, and several additional spectroscopy formats. The operations the model can perform on these include structure prediction, binding site localization, quantitative property analysis, and structure-to-function inference. For a drug discovery lab that currently runs a separate specialized tool for each of those tasks, with separate Python environments and integration engineering between each of them, the ability to unify them in a single model with a one-million-token context represents a workflow consolidation that goes beyond simple performance improvement.

The model card also describes an iterative verification loop that the lab says distinguishes Sage S2.1 Flash from earlier scientific models. The model can call specialized scientific software, observe intermediate numerical outputs and visual feedback from those calls, and revise its subsequent analysis steps based on what it finds. That loop brings its behavior closer to what a computational chemist does during a multi-hour simulation run than what a standard language model produces in a single auto-regressive forward pass. Whether that description maps accurately to what the deployed model does in practice is a question independent testing will need to answer.

The Architecture Traces Back to Z.ai's GLM-5.3-Flash, Released in August

The quiet nature of the release has pushed observers toward the model's configuration files for answers about what is actually inside. An OrcaRouter analysis of those files, published November 16, found that Sage S2.1 Flash uses a model type tagged as `glm5_next`, and that the deployment guide directs users to follow Z.ai's GLM-5.3-Flash serving recipes verbatim, including the need to pass `--trust-remote-code` and apply a specific vLLM patch. Z.ai, the company behind the Zhipu AI GLM series, released GLM-5.3-Flash in August 2026 as a 320B/18B MoE with similar layering. The parameter counts, architecture depth, and expert routing match closely enough that the consensus reading is that Sage S2.1 Flash is built on the GLM-5.3-Flash base with deep scientific fine-tuning on top.

The architecture is specific: 45 layers total, of which 34 use linear attention and 11 use conventional sparse attention. The expert layer deploys 288 routed experts plus one shared expert, with 8 selected per token at inference time. That expert routing design is consistent with both the low active-parameter count and the benchmark pattern, where the model performs very differently across scientific versus general coding tasks, suggesting that domain-specific expert groups have been strongly activated by the scientific fine-tuning process.

Building on an existing base model is not a critique. Most of the most capable specialized models in production pharmaceutical research pipelines today are domain-specific extensions of general-purpose base models, often trained by different organizations than those that produced the base. What matters for the scientific use case is whether the SciPerception layer and the fine-tuning genuinely extend what the base model can do with molecular and protein data beyond what GLM-5.3-Flash already provides. The BabyVision benchmark, a visual scientific reasoning test on which Sage S2.1 Flash scores 75.8 against GLM-5.3-Flash's 51.3, is the most direct evidence on that question. A 24-point gap on a task that tests the specific scientific visual input capability that SciPerception is meant to add is the right benchmark to watch when independent results arrive.

The ModelScope X post published November 16 was the first official communication from any lab channel about the release, appearing more than 24 hours after the weights were already downloadable. It confirmed the MIT license and described the multi-step scientific workflow capability, though it did not explain why the lab bypassed its standard announcement process.

DeepSeek Still Wins on Code, and the Gaps Explain the Model's Design Tradeoffs

The same benchmark table that shows Sage S2.1 Flash leading on scientific tasks shows it trailing on software engineering work. On SWE-bench Pro, the most demanding measure of practical software engineering capability, Sage S2.1 Flash scores 60.6 against DeepSeek-V4.1-Flash-1017's 64.7. The AI Weekly analysis of the release notes that this gap is consistent across the other coding benchmarks in the comparison set, suggesting it reflects a genuine tradeoff in the fine-tuning approach rather than a measurement anomaly.

That tradeoff is predictable from the design. A fine-tuning run that heavily weights scientific reasoning tasks, molecular encoding objectives, and research workflow completion will shift the model's weights toward those domains. The question is whether the base model's coding capability, which GLM-5.3-Flash built in, was partially displaced by the scientific focus, or whether the scientific capability was added on top without eroding general engineering performance. The 4-point SWE-bench gap suggests some displacement occurred. For the target use case, that is acceptable, since a computational chemistry lab running protein structure prediction does not need its AI to also compete with Devin on software engineering leaderboards.

DeepSeek, Qwen, Z.ai, and Mistral all have strong MoE models at similar scale, but none has published MIT-licensed weights with native scientific multimodal input support and a 1M context window in the same package. The open-weight scientific AI space at the 300B-plus scale was effectively empty before this release. The closest precedents are specialized protein structure models like those in the AlphaFold lineage, which are powerful within their narrow domain but cannot read a paper, write analysis code, interpret spectroscopy output, and run iterative simulations within a single unified context.

The Risk Is That Self-Reported Benchmarks From an Unannounced Drop Carry No External Weight

The bear case here is straightforward. Every benchmark result in the model card is InternLM's own measurement, run on InternLM's own harness, with no published run logs and no third-party replication. The comparison set is also missing several models that would appear in any comprehensive evaluation published by an independent benchmark organization, including Claude Haiku 5.5 and Gemini 3.8 Flash, both of which perform competitively on agentic research tasks. That omission does not prove cherry-picking, but it makes the table harder to interpret, since the strongest possible alternative is not present.

The quiet release pattern itself is evidence worth taking seriously. A model that is production-ready and correctly benchmarked at top-of-class performance is almost always announced. The absence of a blog post, a safety evaluation card, and any pricing information at launch suggests one of several possibilities: the lab published the weights as a research artifact rather than a deployment, the model is still being evaluated internally and the Hugging Face upload was premature, or the release is part of a staged rollout timed to a formal announcement that has not yet appeared. None of those explanations undermine the core facts. The weights exist, they are MIT-licensed, and they are downloadable by anyone. But they should calibrate how much confidence to place in the benchmark table before independent results arrive, which will likely take two to three weeks given the size of the model and the compute required to run full benchmark sweeps.

There is also a dual-use consideration that the absence of safety documentation makes harder to evaluate. A model that processes protein sequences and biological molecular data at this scale, in an open-weight MIT-licensed format, sits inside the envelope of biosecurity policy discussions that have been active at several major AI safety organizations. Whether Sage S2.1 Flash was evaluated for misuse potential before release, and what the lab concluded, is unknown from the public record.

What MIT Licensing at 320B Scale Changes for Pharma and Climate Research

The practical question for pharmaceutical companies, climate modeling organizations, materials science groups, and genomics firms is not whether Sage S2.1 Flash tops a leaderboard. It is whether they can run it, modify it, and deploy it in workflows that involve proprietary data, proprietary fine-tuning, and commercial products built on the results. The MIT license answers all three of those questions affirmatively. There are no restrictions on commercial use, no requirement to publish modifications or derivatives, and no research-only clauses that would block deployment in a commercial drug discovery pipeline.

At 645 GB in BF16 or approximately 320 GB in FP8, the model requires either a small multi-node cluster or a single node with 8 or more H100 or H200 GPUs. That is not a small-lab deployment, but it is well within the infrastructure footprint of any pharmaceutical company running serious computational chemistry, any national climate modeling center, or any genomics firm operating at clinical scale. For those organizations, a 1M-context scientific AI that can read a full dataset alongside its own code, reason over molecular inputs natively, and iterate on the analysis in a verified loop represents a different class of automation than the 7B or 13B open-weight models that most research workflows have been built around for the past three years.

The capability to watch most closely for near-term deployment is binding site localization on protein structures. The existing open-source tools for that task require separate geometric deep learning environments, extensive data preprocessing pipelines, and engineering work to connect their outputs to downstream analysis. If Sage S2.1 Flash can perform binding site localization reliably through natural language prompting with native 3D protein structure inputs, it removes an integration layer that currently requires months of engineering. That is an if requiring independent validation, but it is the right if to test, and the benchmark numbers, if they hold, suggest it is not implausible.

Three things to watch in the coming weeks. First, independent benchmark reproductions from university groups and third-party benchmark organizations will arrive within two to three weeks given how quickly the community evaluates newly released weights at this scale. Second, any communication from Shanghai AI Laboratory or Z.ai clarifying the relationship between Sage S2.1 Flash and GLM-5.3-Flash will determine whether this is a collaboration, a licensed build, or simply a case of the deployment guide being more transparent than the announcement. Third, biosecurity and safety organizations will likely weigh in on the dual-use question for a model with native biological sequence processing, since that assessment affects whether and how the broader research community deploys the weights.

The largest open-weight scientific AI on record just appeared on Hugging Face with no announcement, MIT-licensed weights, and benchmark claims that, if they survive independent testing, make the line between open and proprietary research AI look thinner than it has in years.


Key Takeaways

  • 320B total parameters, 18B active per token in a mixture-of-experts architecture under MIT license, making Sage S2.1 Flash the largest open-weight scientific AI model published to date
  • 90.3 on MLE-bench Lite, ahead of DeepSeek-V4.1-Flash-1017 at 89.2, with first place on 4 of 15 benchmarks and second on 7, per InternLM's own evaluations with no independent reproductions yet
  • 10-plus scientific input types including protein sequences, molecular graphs, 3D protein structures, crystal structures, and DNA sequences, unified through the Sage SciPerception architecture alongside natural language
  • 1,048,576-token context window and 645 GB of BF16 weights across 120 shards, with an FP8 variant available, requiring a multi-GPU node with H100 or H200 class hardware
  • No safety documentation and no independent benchmark reproductions at launch, with architecture traces pointing to Z.ai's GLM-5.3-Flash base model from August 2026

Questions Worth Asking

  1. If the GLM-5.3-Flash architecture is the base, how much of Sage S2.1 Flash's scientific edge comes from the SciPerception fine-tuning specifically, versus capabilities already embedded in the base model that simply weren't tested against scientific benchmarks before?
  2. What does it signal about the trajectory of open-source scientific AI that the first 300B-plus model with native molecular and protein inputs arrived with no announcement, no safety card, and no pricing, rather than through a lab's formal product channel?
  3. If independent results confirm the binding site localization and protein structure capabilities, which research software vendors and specialized scientific AI companies face the most displacement from a free MIT-licensed alternative at this performance level?

Sources

  1. Hugging Face model card huggingface.co
  2. Sage S2.1 Flash technical report arxiv.org
  3. OrcaRouter analysis of those files orcarouter.ai
  4. ModelScope X post published November 16 x.com
  5. AI Weekly analysis of the release aiweekly.co

Read Next

Bambu Lab Prices Three 3D Printers From $749 to $2,399

Microsoft MAI Code 2 Claims 96% of Claude Opus at 5% Cost

Nvidia CUDA-Q Cuts Quantum Error Decoding to 45 Minutes

Samsung's Android XR Glasses Sell Out in Four Days

Newsletter

Enjoyed this analysis? Join the list for the email edition.

Daily AI signals. No noise. Built for founders, investors, and operators.

Share:XLinkedIn
</> Embed this article

Copy the iframe code below to embed on your site:

<iframe src="https://techfastforward.com/embed/internlm-sage-21-tops-mle-bench-at-903-with-mit-license" width="480" height="260" frameborder="0" style="border-radius:16px;max-width:100%;" loading="lazy"></iframe>