A year after a Chinese startup briefly rewrote the assumptions of the AI industry, DeepSeek has returned with two new open-source models — V4-Pro and V4-Flash — released on the same day OpenAI unveiled GPT-5.5, in a moment that crystallizes how thoroughly the race for artificial intelligence has become inseparable from the broader contest between nations. The models claim leadership among open-source systems in coding and mathematics, run on Chinese-made chips rather than American ones, and arrive with a rare admission: that the frontier still lies a few months ahead. In that candor, and in th
DeepSeek launches V4-Pro and V4-Flash, challenging frontier models one year after R1 shock
The gap represents approximately three to six months of developmental trajectory.
Mark
Why does the timing matter so much here? It's just a release date.
Mimi
Because R1 landed on January 2025 and shifted how the entire industry thought about the cost of frontier AI. A year later, DeepSeek is saying: we're back, and we're doing it again. That's a statement.
Luke
But we should be careful about what we're actually comparing. R1's shock was partly about the price tag—under $6 million, though some analysts disputed that figure. V4 is a different claim entirely. It's not claiming to be cheaper; it's claiming to be competitive with GPT-5.4 while being open-source.
Mimi
Right. Open-source changes everything. Developers can use it, modify it, run it themselves. That's not just a technical release; it's a distribution strategy.
Mark
And the Huawei chip angle—is that real or symbolic?
Luke
It's both, but we need to be honest about what we know. DeepSeek optimized for Huawei hardware instead of Nvidia. That's confirmed. Whether it actually runs *better* on Huawei chips than Nvidia ones—that's still being tested. The geopolitical significance is real, but the technical validation isn't done yet.
Mimi
The US export restrictions on advanced chips to China have been in place for years. If DeepSeek can run frontier models on domestic hardware, that's a crack in that wall. But Luke's right—we're looking at a preview, not a finished product.
Mark
So what do we actually know versus what we're waiting to find out?
Luke
We know DeepSeek released two models with a 1-million-token context window. We know they claim top performance in coding and math among open models. We know they optimized for Huawei chips. We don't know yet if those performance claims hold up under independent testing. We don't know if the Huawei optimization actually delivers a meaningful advantage. We're waiting on benchmarking.
Mimi
And that's the story—not the release itself, but whether it holds up.
O Pulso
- DeepSeek dropped V4-Pro and V4-Flash on Hugging Face the same morning OpenAI released GPT-5.5, turning a product launch into a symbolic confrontation between two AI powers.
- The models boast a 1-million-token context window — large enough to ingest an entire codebase in a single prompt — pushing the practical ceiling of what open-source AI can handle.
- Rather than claiming parity with the frontier, DeepSeek published its own gap estimate of three to six months behind GPT-5.4, an unusual move that reads as either intellectual honesty or a calculated hedge before independent benchmarks arrive.
- V4 was optimized for Huawei and Cambricon chips instead of Nvidia hardware, making the release a live stress test of China's domestic AI supply chain under US export restrictions.
- Independent evaluators are expected to validate or challenge DeepSeek's benchmark claims within days — the same process that confirmed R1's credibility a year ago, and the same fire these models must now walk through.
A year after a Chinese startup briefly rewrote the assumptions of the AI industry, DeepSeek has returned with two new open-source models — V4-Pro and V4-Flash — released on the same day OpenAI unveiled GPT-5.5, in a moment that crystallizes how thoroughly the race for artificial intelligence has become inseparable from the broader contest between nations. The models claim leadership among open-source systems in coding and mathematics, run on Chinese-made chips rather than American ones, and arrive with a rare admission: that the frontier still lies a few months ahead. In that candor, and in the architecture beneath it, lies a question the industry will spend the coming days trying to answer.
DeepSeek, the Hangzhou startup whose R1 model wiped roughly $600 billion from Nvidia's market value in a single day last January, released two new flagship systems on Friday: V4-Pro and V4-Flash, both published as open-source code on Hugging Face. The timing was pointed — almost exactly one year after R1's debut, and the same day OpenAI released GPT-5.5 into a market that now includes a trillion-dollar Anthropic and an AI competition that has hardened into explicit trade policy.
Both variants use what DeepSeek calls a Hybrid Attention Architecture, designed to maintain coherence across very long exchanges. The result is a 1-million-token context window — enough to process an entire codebase or a book-length document in one prompt. V4-Flash emphasizes speed and cost; V4-Pro chases peak performance.
By DeepSeek's own benchmarks, V4-Pro leads all open-source models in coding and mathematics, and trails only Google's closed-source Gemini 3.1-Pro in world knowledge. Against the current frontier, DeepSeek says the gap is real but quantifiable: roughly three to six months of development. Publishing that estimate rather than claiming parity is unusual in an industry that typically announces only favorable comparisons — whether it reflects genuine transparency or strategic expectation-setting, independent testing will soon reveal.
The chip dimension carries its own significance. DeepSeek optimized V4 for Huawei and Cambricon hardware, bypassing Nvidia and AMD entirely — a reversal of standard practice. Running a model of this scale on Huawei's Ascend chips would serve as a meaningful proof of concept for China's domestic AI supply chain, which has operated under US export restrictions since 2022. V4 doesn't dissolve that constraint, but it tests it in a commercially visible way.
Both releases are previews, not final versions, and DeepSeek's self-reported benchmarks remain unconfirmed. R1's claims were validated by external testing within days. Whether V4 survives the same scrutiny will be known before the week is out.
DeepSeek, the Hangzhou startup that rattled Silicon Valley a year ago with its R1 model, released preview versions of two new flagship systems on Friday: V4-Pro and V4-Flash, both posted to Hugging Face as open-source code. The timing was deliberate—almost exactly one year after R1's January 2025 debut, which erased roughly $600 billion from Nvidia's market value in a single trading day by suggesting that frontier-class AI could be built for far less compute than the industry had assumed.
The new models arrive into a fiercer landscape. OpenAI released GPT-5.5 the same day. Anthropic is valued at $1 trillion on secondary markets. The US-China AI competition has hardened into explicit trade and technology policy. DeepSeek's second act is landing in contested territory.
Both V4 variants employ what DeepSeek calls a Hybrid Attention Architecture, a technique designed to let the model retain context across long conversations without degrading in quality as the conversation lengthens. The practical upshot is a 1-million-token context window—enough to process an entire codebase or a book-length document in a single prompt. The Flash variant prioritizes speed and cost efficiency; the Pro variant chases peak capability.
On DeepSeek's own benchmarks, V4-Pro leads all open-source models in coding and mathematics. In world knowledge, it trails only Google's closed-source Gemini 3.1-Pro. Against the current frontier—OpenAI's GPT-5.4 and Gemini 3.1-Pro—DeepSeek says V4-Pro falls only marginally short, and offers an unusual candor: the gap represents approximately three to six months of developmental trajectory. That framing is striking. AI model releases typically emphasize comparisons where the new system leads. That DeepSeek published a gap estimate rather than claiming parity suggests either unusual intellectual honesty or a strategic move to set conservative expectations before independent evaluation arrives.
The chip story carries geopolitical weight. DeepSeek optimized V4 for hardware from Chinese chipmakers Huawei and Cambricon, according to reporting from Reuters citing The Information. The company did not give Nvidia or AMD early access for optimization—a reversal of standard industry practice, where Western chipmakers typically receive new model weights first for performance tuning. Running a frontier-class model at this scale on Huawei's Ascend chips rather than Nvidia H100s or H200s would constitute a significant proof of concept for China's domestic AI hardware supply chain, which has operated under US export restrictions since October 2022. The V4 release does not dissolve that geopolitical constraint, but it tests its boundaries in a commercially visible way.
Both models are preview releases, not final production versions. Independent benchmarking has not yet been completed. DeepSeek's own benchmarks should be treated as preliminary until third-party evaluation confirms them. The same caveat applied to R1, whose claims were broadly validated by external testing within days of release. Whether V4 holds up to the same scrutiny will be clear within the week.
Citações Notáveis
DeepSeek says V4-Pro falls only marginally short of GPT-5.4 and Gemini 3.1-Pro, with a developmental gap of approximately three to six months.— DeepSeek's self-assessment in model benchmarking
Quer a matéria completa?
Leia o original em TNW ↗
Technology