DeepSeek’s Update Brings R1 Model Back to the Top

DeepSeek R1 was praised for its performance and relatively lower price compared to US rivals like OpenAI and Google Gemini when it launched last January.

The release triggered shockwaves across the tech market, with most shares outside China plummeting. The frenzy wiped out billions from Nvidia’s market cap and even caused OpenAI to slash prices.

While the R1 euphoria later grew thin, the tech startup is back with an updated R1 reasoning model dubbed DeepSeek-R1-0528. The new update is available for commercial use under the MIT license, with DeepSeek claiming it has an improved benchmark performance, better front-end capabilities, and fewer hallucinations than the original R1 model.

As usual, there are opinions suggesting that the DeepSeek-R1-0528 update can rival Google’s Gemini and OpenAI’s o3 in many areas. To get a glimpse, Techopedia has rounded up everything you need to know about the DeepSeek-R1-0528 and what it means for competitions in the AI race.

Table of Contents Table of Contents

1. Key Takeaways
2. Deepseek Latest Update: Minor Change With Major Impact
3. Where Does DeepSeek Model Rank Against Leading LLMs?
4. The Bottom Line
5. FAQs
6. References

Show Full Guide

Table of Contents

1. Key Takeaways
2. Deepseek Latest Update: Minor Change With Major Impact
3. Where Does DeepSeek Model Rank Against Leading LLMs?
Show Full Guide
4. The Bottom Line
5. FAQs
6. References

Deepseek Latest Update: Minor Change With Major Impact

DeepSeek confirmed through popular model hub Hugging Face that its DeepSeek R1 model has undergone a minor version upgrade, with the current version being DeepSeek-R1-0528.

The new model features approximately 685 billion parameters, with 37 billion active per token, supports a 128K token context window, and delivers advanced reasoning capabilities, achieving performance comparable to OpenAI’s o3 and Google’s Gemini 2.5 Pro.

Performance metrics show R1-0528 improved LiveCodeBench scores from 63.5% to 73.3%, and on the AIME 2025 mathematics test, accuracy jumped from 70% to 87.5%.

This result means R1-0528 has a better performance on tasks that require multi-step logic, contextual understanding, and precise inference.

Also, in benchmarks like MMLU, GSM8K, BBH, and HumanEval, DeepSeek R1-0528 achieved a median score of 69.45. This is largely perceived to be a huge feat for an AI startup with no Silicon Valley backing.

The Chinese AI startup also revealed a distilled model, DeepSeek-R1-0528-Qwen3-8B, built on Alibaba’s Qwen3-8B. DeepSeek reports that the model outperforms Qwen3-8B by +10.0% on the AIME 2024 benchmark and matches the performance of Qwen3-235B-thinking among open-source models.

Beyond having capabilities for complex math reasoning, DeepSeek also revealed on X that R1-0528 offers a lower hallucination rate, better support for function calling, and better experience for vibe coding, which now rivals outputs from commercial models like OpenAI o3 and Claude 4.

While there is some positive feedback across many Reddit discussion forums on R1-0528, there are still users who think DeepSeek hasn’t solved a major challenge. One user, for instance, voiced concerns, saying that the model “needs single conversation memory first,” criticizing DeepSeek’s “horrific memory” as its biggest weakness.

Where Does DeepSeek Model Rank Against Leading LLMs?

According to a post on X by Artificial Analysis, a benchmarking group that compares LLM API performance, the “gap between open and closed models is smaller than ever.” Their recent evaluation ranks DeepSeek R1-0528 as the world’s second most capable AI model, delivering performance equal to leading proprietary AI models and trailing only OpenAI’s o3.

The group further noted that DeepSeek has outpaced Meta’s LLaMA models, xAI’s Grok, and Anthropic’s Claude models, calling it the “undisputed open-weights leader.”

Artificial Analysis said:

“Open-weights models have continued to maintain intelligence gains in line with proprietary models. DeepSeek’s R1 release in January was the first time an open-weights model achieved the number 2 position, and DeepSeek’s R1 update today brings it back to the same position.”

What’s even more disruptive is DeepSeek’s inference cost. Some estimates suggest that its latest R1-0528 model, running at full 128k context, can generate tokens for just a few dollars per million, making it 5 to 7 times cheaper than top-tier proprietary models like Gemini 2.5 Pro or Claude 4.

This advantage in price, combined with solid reasoning and competitive benchmark scores, makes DeepSeek very attractive for startups, academic research, and enterprise applications operating at scale.

deepseek-r1-0528 is now competing head-to-head with o3 and gemini 2.5 pro

and it’s open-source.

this release makes one thing clear: deepseek is not catching up, it’s competing pic.twitter.com/3fFq5uVG6n

— Haider. (@slow_developer) May 29, 2025

The Bottom Line

DeepSeek’s R1-0528 update shows just how quickly the AI landscape is shifting. It doesn’t reinvent the model, but the improvements are targeted and meaningful.

Worthy of note is that it does all this while keeping inference costs low and remaining open-source. That’s not a small detail in a market where price and access can limit experimentation.

The benchmarks matter, but so does the model’s positioning: R1-0528 now sits just behind OpenAI o3, and ahead of many proprietary tools. There are still gaps, especially around memory, but DeepSeek’s progress is difficult to ignore.

FAQs