DeepSeek rocked Silicon Valley in January 2025 – one year on and it looks set to shake things up again with a powerful new model release

When DeepSeek hit the scene in January 2025, the Chinese AI company sent shockwaves through the industry.

The launch of the company’s R1 model, a fine-tuned version of its previous V3 range, prompted a market meltdown. Broadcom and Nvidia shares nosedived, wiping $593 billion from the latter’s market value alone.

DeepSeek was a big wake-up call for US counterparts for a number of reasons. Chief among them was the fact it proved Chinese AI developers weren’t as far behind as some industry stakeholders believed – and the fact its advances were open source.

Notably, the most jarring aspect of the DeepSeek R1 launch lay in the reputed cost of building it. Built on a shoestring budget compared to those offered by US counterparts, this raised serious questions about the huge financial outlay associated with AI training.

In the year prior, big tech talking heads repeatedly touted the prospect of unfathomably large model training costs. Speaking in July 2024, Anthropic CEO Dario Amodei predicted skyrocketing fees in this domain, suggesting that future models could cost upwards of $100 million to build.

It’s safe to say that DeepSeek poured cold water on these bold claims and, in doing it with such bravado, created a sense of unease among US rivals. R1 went toe-to-toe with anything US providers could offer in terms of performance.

The model was trained on 37 billion active parameters and 671 billion total parameters, according to figures released at the time.

For context, benchmarks published by the Chinese firm showed it outpaced Anthropic’s Claude 3.5 and OpenAI’s GPT-4o in key areas, particularly in coding capabilities.

DeepSeek was an instant hit, with users globally flocking to the new application. Free to access, the DeepSeek app toppled ChatGPT as the top free application on the Apple App Store.

But one year on from the launch, where does DeepSeek stand? The initial furor and excitement has subsided to some extent, but the company is still plugging away – and so are users.

Figures from OpenRouter show token consumption for DeepSeek V3 0324 topped 7.27 trillion across 2025, ranking it fifth behind counterparts such as Claude Sonnet 4 and Gemini 2.0 Flash.

DeepSeek has been quietly innovating

DeepSeek’s V3 foundation model has had a number of updates across the last year, with V3.1 arriving in August. A hybrid open weight model, this marked a first major step toward agentic reasoning, combining thinking and non-thinking modes within a single system.

Once again, performance and cost were key talking points, performing well on coding benchmarks while offering more efficient compute costs compared to other models on the market.

Hot on the heels of V3.1, the Chinese firm launched V3.2 and V3.2-Speciale in December. The first of these, according to DeepSeek, places a strong focus on mimicking human reasoning while the Speciale range is designed primarily for mathematical contexts, boasting "maxed-out reasoning capabilities".

Both models offered equivalent performance to OpenAI’s GPT-5, with the team describing it as “your daily driver” in a clear nod to its use in a similar fashion to ChatGPT.

Let’s talk about cybersecurity

Wherever a conversation about DeepSeek arises, questions over security and data privacy haven’t been far away. Indeed, the R1 launch prompted an array of warnings from industry stakeholders about the risks associated with the model.

Indeed, February 2025 saw Australian lawmakers ban the application from government devices and systems, citing security concerns. Since then, lingering worries about the Chinese AI provider have persisted across both government and industry.

Speaking to ITPro in August last year, Andy Ward, SVP International at Absolute Security, said enterprises should exercise extreme caution when using the application, describing it as akin to “printing out and handing over confidential information”.

Research on the potential risks associated with DeepSeek have been highlighted in a slew of studies since its release. Less than a month after launch, researchers at Cisco warned the model contained “critical safety flaws” that left it open to an array of jailbreak techniques that could be pounced on by threat actors or leave enterprises vulnerable to data exposure.

Separate research from Harmonic Security found DeepSeek displays a disproportionately high risk of sensitive data exposure compared to other models out there on the market - even Chinese offerings.

Analysis by the company found the use of Chinese AI models is growing, with DeepSeek accounting for 25% of overall usage alongside offerings such as Alibaba’s Qwen range.

Crucially, however, DeepSeek accounts for 55% of sensitive exposure, which Harmonic researchers said was partly due to its popularity with coders.

DeepSeek V4 could shake things up again

The arrival of DeepSeek V4 appears to be imminent, according to recent speculation.

Reports from The Information, citing sources with intimate knowledge of the development process, said this latest iteration could arrive in early 2026.

This model will place a strong focus on code-generation features, a recurring theme in recent months amidst the arms race between Anthropic, OpenAI, and Google.

According to the publication, early benchmarks suggest the model will outperform the Claude and GPT ranges on this front.

As with the launch of DeepSeek R1, the Chinese firm is also expected to confirm another technical breakthrough on model development in terms of contending with lengthy code prompts.

If this speculation comes to fruition, we could see another scramble among US AI developers to one-up the company.

FOLLOW US ON SOCIAL MEDIA

Make sure to follow ITPro on Google News to keep tabs on all our latest news, analysis, and reviews.

You can also follow ITPro on LinkedIn, X, Facebook, and BlueSky.