Checking in With DeepSeek – Banyan Hill Publishing

Back in January, DeepSeek shocked the world when it dropped a frontier-scale AI model for a fraction of the cost of its American rivals.

The release of the DeepSeek-R1 proved that China could punch above its weight in high-level reasoning.

And as I mentioned back then, it also changed the trajectory of the AI race.

It was a clear sign that Beijing wanted to close the gap with the United States, and it proved that China was not slowing down.

But I saw it as a good thing. And I believe I’ve been vindicated. Because it finally pushed U.S. policymakers to treat artificial intelligence as a national priority.

I’m convinced it’s one of the reasons the White House recently created a new cross-agency AI development plan called the Genesis Mission that could represent a Manhattan Project for AI.

And it certainly was a factor in the private sector pouring billions of dollars into new training clusters this year.

A move that seems to be paying off.

ChatGPT-5 arrived this year with top scores in long-context reasoning. Google recently launched Gemini 3 and advanced multimodal performance even further. And Anthropic’s Claude has stealthily become the leader of the enterprise AI race.

But that doesn’t mean DeepSeek has been sitting still.

Last week, the company resurfaced with a new release called DeepSeek V3.2 and V3.2 Speciale.

The announcement didn’t shock the world like DeepSeek’s January release, but the details are still eye-opening.

Because if the numbers DeepSeek published are accurate, then China just delivered its strongest open-weight challenger yet.

Which makes this the perfect time to check in with DeepSeek.

New Benchmark Claims

DeepSeek says its V3.2 Speciale model earned gold-level performance on four high-end academic benchmarks. These include the 2025 International Mathematical Olympiad (IMO), the China Mathematical Olympiad (CMO), the International Olympiad in Informatics (IOI) and the ICPC World Finals.

Obviously, these aren’t simple tests.

They are the hardest math and coding challenges in the world, and they are usually dominated by elite research labs. American teams often post strong results, but they rarely release open-weight models that score at the very top.

DeepSeek claims it has now done exactly that.

The company also disclosed something unusual in its technical report. It said the model uses a system called DeepSeek Sparse Attention to handle long-context problems more efficiently.

It also said that more than 10% of its total compute budget was spent on reinforcement learning for reasoning and agentic behavior. That’s unusually high for an open-weight model. If true, it would help explain why DeepSeek is framing V3.2 as a “reasoning-first” model instead of a general-purpose chatbot.

Here is how the company says it stacks up.

As you can see, DeepSeek’s new models appear to match or come close to the top scores posted by GPT-5 and Gemini 3 on narrow reasoning tasks like math and structured problem solving.

These numbers are impressive, but they come with an important caveat.

They haven’t been independently audited. And until they are, we need to treat them as promising claims rather than confirmed breakthroughs.

However, there are parts of this release we can confirm.

The weights are available online, and developers have already begun running local inference tests. Early users say the model handles multi-step reasoning better than earlier DeepSeek versions. And the sparse attention mechanism seems to be real based on the published code.

But the picture becomes less clear when we step beyond the math and coding scores.

A few independent groups, including a research team that collaborates with NIST, tested earlier DeepSeek models this year. Their conclusion was that those versions still lag behind the best American systems in broad knowledge, tool use and real-world reliability.

These findings do not contradict DeepSeek’s new numbers, but they do underscore something important.

Scoring well on math contests doesn’t guarantee general intelligence. It simply shows strength in one part of the larger puzzle.

But general intelligence is what counts in the long run.

This is the same gap we talked about in January. Right now, U.S. companies still hold the lead in scaled multimodal training, global safety testing and integrated platform deployment.

OpenAI has the best tool-use system in production. Google has the most developed memory architecture. Anthropic has the strongest track record on reliability and reasoning stability. And together, these companies have access to the largest training clusters on the planet.

DeepSeek is still chasing these companies. But that doesn’t mean the gap remains as wide as it once was.

DeepSeek’s new model is advancing at a pace that would have seemed unrealistic just a year ago. And the fact that it can ship open-weight models with near-frontier math scores should worry anyone who thinks the United States can afford to coast.

Because every time China advances in AI, it puts pressure on the United States to move even faster.

Here’s My Take

DeepSeek claims to have trained V3.2 using more than 1,800 synthetic environments and more than 85,000 tool-use prompts. These include search tasks, coding tasks and multi-step agent tasks.

Agentic behavior is the next major frontier in AI. Models that can reason, plan and take actions on their own will shape everything from software development to national security.

That’s why I’ll continue to keep a close eye on DeepSeek.

Because the company says it will continue scaling its agentic pipeline. And if it stays on this trajectory, we should expect even more ambitious models in 2026.

This means the United States has to keep pushing its own pace.

We still have the strongest AI companies in the world. But this release sends a clear message that the race to artificial superintelligence (ASI) is closer today than it was in January.

And both sides know it.

Regards,

Ian KingChief Strategist, Banyan Hill Publishing

Editor’s Note: We’d love to hear from you!

If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to [email protected].

Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!