DeepSeek R1-0528 Just Dropped. Even the Name Feels Like a Puzzle

DeepSeek has released a new open-source model, R1-0528, trained entirely from scratch. It’s a full-stack code model, positioned as a major upgrade over its earlier versions. The company claims strong performance on several industry benchmarks.

What DeepSeek Is Saying

DeepSeek describes R1-0528 as a new version of its base model, now publicly available on Hugging Face and GitHub. It was trained from scratch on 6T tokens using a mixture of English, Chinese, and 87% code.

"We trained it entirely from scratch, using our own data and infrastructure, to produce stronger reasoning and coding performance."
- DeepSeek Labs, May 2025

They report improvements across multiple benchmarks compared to their previous model R1, including AIME, LiveCodeBench, and GPQA.

🧠 What That Means (In Human Words)

This new model update - R1-0528 - shows big improvements across key reasoning and code-generation tasks.
It outperformed models like Grok 3 Mini and Alibaba’s Qwen 3 in coding tasks and showed stronger multilingual and math skills than its earlier version.

Here’s what it nailed:

• Code Generation: 73.3% pass@1 on LiveCodeBench (up from 63.5%)

• Math Reasoning: 87.5% on AIME problems

• Multilingual Coding: 71.6% accuracy (up from 53.3%)

• GPQA Reasoning: 81% accuracy

• Humanity’s Last Exam: Doubled performance (from 8.5% to 17.7%)

But What Does All of That Mean?

Yes, this is hard. Everyone is saying the same thing - that their new model is better than the last.

And on paper, they all are.

Because the bare minimum for a release today is that it performs better on benchmarks.

Let’s try to make sense of what we’re actually comparing.

Right now, we’ve mostly seen two types of benchmarks:

1. Hands-on - things like SWE-bench and LiveCodeBench. These simulate real-world programming tasks.

2. Academic - things like AIME, GPQA, MATH. These are about logic, puzzles, and conceptual reasoning.

One came to work, the other came to play chess.

DeepSeek R1-0528 is a big step up over its last version.

But there’s no SWE-bench score published. And that’s the benchmark used by GPT-4.1 and Claude Opus to show their real-world strength.

So can we say DeepSeek beats GPT or Claude?

No. Not yet.

We just don’t have the same test results to compare.

We made a table but it did not help :)

Bottom Line

• Model: DeepSeek R1-0528

• Access: Open source, available on Hugging Face and GitHub

• Best For: Coding, AI tinkering, experimentation

• Benchmarks: Strong in academic reasoning and hands-on code generation

• Should You Try It? Yes, if you're curious about where open-source coding models are heading next

• Cost: Free to use

Frozen Light Team Perspective

This is a classic case of not finding the information you actually need to solve your dilemma.

If you're a programmer trying to understand what’s better-you’ll just have to try it yourself.

From research we’ve done in GitHub communities, here’s what we can tell you:

When it comes to practical, hands-on usage - moving things, plugging things, getting stuff done - ChatGPT and Claude consistently get higher scores in actual dev environments.

And to be honest, DeepSeek isn’t showing up in many real-world coding conversations yet.

That doesn’t mean it’s bad.

The rest? That’s up to you to try and decide what works best for you.

The rest? That’s up to you.