Intern-S2 and DeepSeek V4 Pro are both open-weight AI models released recently, but they solve fundamentally different problems. Intern-S2, released by InternLM on September 13, 2026, is a 403-billion-parameter model designed specifically for scientific research, while DeepSeek V4 Pro, which became available on August 13, 2026, is a 1.6-trillion-parameter general-purpose model with 49 billion active parameters. The two models were recently compared head-to-head on the same benchmark table, but that comparison reveals more about their differences than their similarities.

What Makes These Models So Different?

The key distinction lies in what each model was trained to do. Intern-S2 was built to read dense scientific literature, reason over molecular and materials data, and work with time-series signals. It includes multimodal capabilities, meaning it can process text, images, and time-series data together. DeepSeek V4 Pro, by contrast, is a text-focused general reasoning model designed to handle a wide range of tasks without specialization.

When researchers ran both models through the same scientific benchmarks, the differences became stark. On biology instruction tasks, Intern-S2 scored 55.71 compared to DeepSeek V4 Pro's 9.14. On biomolecular instruction, Intern-S2 achieved 53.95 versus 12.06. These gaps reflect Intern-S2's specialized training on scientific literature. However, on general knowledge tests like MMLU Pro, a widely used knowledge benchmark, Intern-S2 scored 89.77 while DeepSeek V4 Pro scored 86.86, a much tighter margin.

How Do Their Technical Specifications Compare?

Beyond performance, the two models differ significantly in their technical architecture and capabilities:

• Model Size and Architecture: Intern-S2 uses 403 billion total parameters with a mixture-of-experts design, while DeepSeek V4 Pro uses 1.6 trillion total parameters with 49 billion active parameters, making it a much larger download but potentially more efficient during inference.
• Context Window: Intern-S2 can process up to 256,000 tokens for text reasoning and 64,000 tokens for multimodal work, roughly equivalent to 200,000 and 50,000 words respectively. DeepSeek V4 Pro offers a 1-million-token context window, roughly 750,000 words, and can generate up to 384,000 tokens of output.
• Input Types: Intern-S2 accepts text, images, and time-series data. DeepSeek V4 Pro is text-led and does not process time-series signals or images natively.
• Licensing: Both models are genuinely open-weight. Intern-S2 uses the Apache 2.0 license, while DeepSeek V4 Pro uses the MIT license, both of which permit commercial use and modification without regional restrictions.

What About Pricing and Availability?

One critical difference emerges when considering real-world adoption: pricing transparency. DeepSeek V4 Pro has a published API cost of $0.66 per million input tokens and $1.98 per million output tokens during off-peak hours, roughly doubling during peak times. This is a metered, predictable cost that teams can budget for before committing to the model.

Intern-S2 has no public pricing. Organizations can either self-host the model on their own hardware, which means near-zero marginal cost per token if they already run GPU capacity, or use the official Intern API with an opaque token allocation system. This difference shapes how teams actually adopt these models. DeepSeek V4 Pro is a service you can price before you commit; Intern-S2 is infrastructure or a pilot.

How Should Teams Choose Between Them?

The choice depends entirely on the job at hand. Pick Intern-S2 when the work is scientific: reading figure-dense literature, reasoning over molecular or materials data, working with time-series signals, or building an agent that needs to survive a long research task. The model was trained specifically for these applications, and the benchmark data shows it excels at them.

Choose DeepSeek V4 Pro when you need a general-purpose reasoning model with a massive context window, predictable pricing, and the ability to handle diverse tasks without specialization. DeepSeek V4 Pro also has an independent score from Artificial Analysis, which rates it at 36 on its Intelligence Index with a 69 on the coding sub-index, placing it in the top ten of 113 models evaluated. Intern-S2, released just this morning, has no independent index entry yet, meaning its entire performance story comes from benchmarks published by its own creators.

Both models are genuinely open and permissive in their licensing, so the decision is not about access or control. It is about fit. Intern-S2 is narrower and deeper; DeepSeek V4 Pro is broader and more general. Neither is a universal winner, and that is precisely the point. The open-weight AI landscape is maturing enough that different models can serve different purposes without one dominating the other.