What happened: Intern-S2 is fully open source
In mid-September 2026, Shanghai AI Laboratory released the complete weights of Intern-S2. This was not another marketing move that ships a model while only exposing an API; it was genuine open source, with both the code repository and the weights published at the same time. The model first appeared under the name Intern-S2-Preview-397B at the WAIC conference in July 2026, and around September 13 the formal weights went live, with a clear positioning: to become the open-source multimodal base that understands science best.
Intern-S2 scales to 397B parameters and uses a Mixture-of-Experts (MoE) architecture. One thing must be stated plainly: 397B is a base-level scale, which means it has the capacity to carry general ability, long-horizon scientific reasoning, and multimodal understanding, rather than being a lightweight chat toy. For Chinese technical practitioners, the significance of this open-source move is that you can finally obtain a multimodal base on par with the closed-source first tier, one that is auditable and modifiable, instead of merely queuing behind someone else's API.
The official entry points come in three places: the GitHub repository at https://github.com/InternLM/Intern-S1 (where Intern-S2's README lives), the HuggingFace weights at https://huggingface.co/internlm/Intern-S2-397B, and the ModelScope entry Shanghai_AI_Laboratory/Intern-S2-397B. In addition, there is a free API quota; the online experience address is https://chat.intern-ai.org.cn/, and the quota policy is at https://internlm.intern-ai.org.cn/api/strategy.
A detail easily overlooked: Intern-S2 has been deeply co-optimized with the domestic computing ecosystem, especially Ascend. In the current external environment, this is not merely a performance matter but a realistic supply-chain choice.
What Memory Decoder is: decoupling knowledge from reasoning
If you treat Intern-S2 merely as a larger multimodal model, you miss its most worth-explaining design. The real core is an architectural idea called Memory Decoder.
The traditional approach goes like this: if you want the model to hold knowledge of a professional domain, you either retrain the base (which is extremely costly and prone to catastrophic forgetting) or bolt on a RAG (retrieval-augmented generation) pipeline that stuffs relevant documents into the context at inference time. RAG is useful, but it mixes knowledge and reasoning together. Every inference must retrieve and re-organize context anew, and in long-horizon scientific tasks this overhead and uncertainty are both troublesome.
Memory Decoder takes a different path: it introduces pluggable external professional memory modules and fully decouples knowledge storage from logical reasoning. Put simply, the base is responsible for general reasoning and understanding, while domain knowledge is encapsulated into an independent, replaceable memory module (such as the accompanying Intern-MemDec-4B). When switching domains, you do not need to retrain the base, nor repeatedly retrieve at every inference step; you only need to mount the corresponding memory module.
This design matters especially for long-horizon research. Tasks such as materials, life sciences, and chemical synthesis are typically combinations of multi-step reasoning and professional facts. RAG may drift in retrieval and bloat context at every step; by front-loading professional knowledge into structured memory, the reasoning process can invoke it stably, without hurting general ability or polluting the base. In a sense, Memory Decoder answers a more fundamental question: as knowledge grows exponentially, should we burn knowledge into the model, or externalize it to the model? Intern-S2 chose externalization, and specifically a pluggable, auditable externalization.
There is also a companion Mobius architecture: reverse residual connections plus dynamic latent-space reasoning. According to the official claim, end-to-end inference efficiency improves by nearly four times. Efficiency matters enormously for a 397B-scale model, because no matter how strong a model is, it is empty talk if it cannot run.
Capability profile, with honest caveats
Next is the part everyone cares about most: how strong is it really. Let us say this upfront: the evaluation conclusions below mostly come from the model developer's own reports and must be flagged as official claims pending independent verification.
On general ability, the official side claims Intern-S2 sits in the open-source first tier. On scientific long-horizon tasks such as Biology-Instructions, Mol-Instructions, and MP20 (materials and life-science related), the official side claims it leads a host of open and closed models. On math long-horizon reasoning, specifically IMO-Proof and AdvancedMathBench, the official side claims performance on par with Gemini 3.1 Pro. On coding and agentic directions, on TerminalBench2.1 and SWEBench-Pro, the official side claims it surpasses Kimi-2.7-Code and DeepSeek-V4-Pro, second only to GLM5.2; on SWEBench-Multilingual it ties with GLM5.2.
Let us complete the picture: these numbers are dazzling, but they currently come mostly from the model developer's own evaluation reports. Independent third-party reproduction remains limited. We write them down faithfully here, both to show you the claimed ceiling and to remind you that, before seeing non-official, repeatable, cross-institution re-tests, it is safer to treat these numbers as a potential range rather than established fact.
This is also a judgment with a point of view: evaluating an open-source base should not only look at benchmark peaks, but also at whether it is auditable, reproducible, and modifiable. On auditability, Intern-S2 has a natural advantage because code and weights are both public; but the answer sheet on reproducibility still needs the community and time to write.
Side-by-side: Intern-S2 versus mainstream models
The table below is a qualitative comparison rather than precise benchmarking, intended to help you build a coordinate system quickly. Dimensions include model nature, parameters, general ability, scientific professional ability, domain extension method, long-horizon research, domestic computing adaptation, and usage cost.
| Dimension | Intern-S2 (open) | Gemini 3.1 Pro (closed) | Other open multimodal (Qwen-VL / GLM-4.5V / InternVL) |
|---|---|---|---|
| Model nature | Open (repo plus weights) | Closed (API only) | Open |
| Parameter scale | 397B (MoE) | Not disclosed | Tens of B to over a hundred B |
| General ability | Open first tier (official) | Closed first tier | Medium to upper-medium |
| Scientific ability | Leads materials and life science (official) | Strong | Generally weak |
| Domain extension | Pluggable Memory Decoder | Not externalizable | Mostly retrain or RAG |
| Long-horizon research | Strong (official: near Gemini 3.1 Pro) | Strong | Average |
| Domestic compute fit | Deep Ascend co-optimization | None | Partial |
| Usage cost | Free weights plus free API | Pay per use | Free weights, self-funded compute |
Two points stand out from this table. First, in the niche of open-source and science-oriented models, Intern-S2 currently has almost no rival. Closed models are strong but not auditable, while other open multimodal models are multimodal but weak on scientific long-horizon tasks. Second, its core differentiator is not a larger parameter count but the domain extension method brought by Memory Decoder: others retrain or add RAG to switch domains, whereas it swaps a memory module. This is an architectural distinction worth noting separately.
For practitioners deciding where to invest, the practical takeaway is simple: if your work is scientific and long-horizon, and you value being able to audit and modify the model, Intern-S2 is currently the strongest open option. If you need a turnkey, fully managed service and do not care about auditing weights, a closed model may still fit better. The two are not strictly competitors; they serve different constraints.
Who it helps, and who it does not
Every tool has applicable boundaries, and Intern-S2 is especially clear about this.
For research institutions, university labs, and enterprise teams with some computing reserves, it is a very cost-effective choice: you can download the weights, mount a memory module for your own domain, and obtain a base that understands your field without training from scratch. On professional long-horizon tasks such as materials, biomedicine, and chemistry, this base-plus-professional-memory combination is more controllable than either bolting RAG onto a general model or retraining outright.
For small teams and individual developers, the reality must be sober: the 397B scale is extremely unfriendly to local deployment. Those who can actually run the full model remain a minority of institutions holding large compute. For individuals, the more realistic path is not local deployment but the free API quota, using it as a powerful cloud-based research assistant.
One more caveat to flag: Intern-S2-Preview-397B is scheduled to be taken offline on October 31, 2026, and users of the preview version need to migrate to the formal release. If you are prototyping on the preview, plan the migration now to avoid a broken pipeline later.
There is also a cost angle often ignored: open source does not mean free to run. The weights are free, but the electricity, memory, and operations needed to actually serve a 397B model are real expenses. If a small team only needs to ask occasional professional questions, the free API is usually cheaper than building its own cluster, and it also avoids the headache of version upgrades.
How to get started: three paths and a free quota
There are three clear paths to start with Intern-S2; you do not need complex deployment commands, just get the resources first.
Path one, the GitHub repository. Go to https://github.com/InternLM/Intern-S1, where Intern-S2's README and usage notes live, suitable for developers who want to read, modify, and audit the code.
Path two, HuggingFace weights. Go to https://huggingface.co/internlm/Intern-S2-397B and pull the weights directly, suitable for workflows already connected to the HF ecosystem.
Path three, ModelScope. Search Shanghai_AI_Laboratory/Intern-S2-397B, suitable for the domestic network environment, with a relatively handy download and inference toolchain.
If you do not want to, or temporarily cannot, deploy locally, the easiest method is the free API: try it online at https://chat.intern-ai.org.cn/, and see the quota and application policy at https://internlm.intern-ai.org.cn/api/strategy. Use the API first to validate whether your task scenario fits, then decide whether to invest compute in local deployment. That is usually the more rational order.
One practical note for first-time users: start with the free API before touching weights. It lets you learn the model's strengths and failure modes on your own tasks at zero cost, and the prompt patterns you develop there transfer directly once you later run the weights locally or on a partner cluster. Treat the API as a low-risk evaluation, not a weakest fallback.
Worth repeating: the most defensible reason to pay attention is not any single benchmark, but the combination of open weights, auditable code, and a genuinely different memory architecture. Those three together are what separate a real open-source base from a passing demo.
A final word with a point of view: Intern-S2 is not another product of a parameter race. Its value lies in using Memory Decoder to outsource knowledge while keeping reasoning in-house, and simultaneously putting both code and weights into the community's hands. We should of course remain cautious about official evaluations, but we should also give projects with genuine open source and genuine architectural innovation a bit more serious attention.
FAQ
Q1: How is Intern-S2 related to Intern-S1?
A1: Intern-S2's code repository is hosted at https://github.com/InternLM/Intern-S1, where its README is also published. You can understand it as a continuation of the Intern series under the same open-source organization; S2 is the new-generation multimodal base, while the S1 repository hosts its code and documentation.
Q2: Is this genuinely open source, or just an API?
A2: It is genuinely open source. It provides both the official GitHub repository (with README and usage) and the complete model weights on HuggingFace and ModelScope, not merely a remote API. You can download, audit, and modify the weights and code, which is fundamentally different from closed services that are API-only.
Q3: Can I run it locally?
A3: For the vast majority of individuals and small-to-medium teams, the realistic answer is no. The 397B MoE scale demands extremely high memory and cluster size, and full local deployment requires the resources of large computing institutions. The more feasible option for individuals is the free API quota at https://chat.intern-ai.org.cn/, treating it as a cloud-based research assistant.
Q4: How do I use the Memory Decoder memory modules?
A4: The core idea is to mount pluggable external professional memory modules (such as the accompanying Intern-MemDec-4B), decoupling domain knowledge from base reasoning. The specific application, training, and mounting methods follow the official GitHub README; for now, rely on official documentation and community practice, and plug in your professional-domain memory as needed.
Q5: Are those numbers like "on par with Gemini 3.1 Pro" trustworthy?
A5: Treat them with caution. These conclusions mainly come from the model developer's own evaluation reports (official claims), and independent third-party reproduction remains limited. We have faithfully flagged them in the text as official claims pending independent verification. Before seeing repeatable, cross-institution, non-official re-tests, it is wiser to treat these numbers as a potential range rather than established fact.