DeepSeek V4 Preview Drops: China's AI Challenger Returns with Million-Word Memory

DeepSeek dropped preview versions of its V4 model on Friday, exactly one year after its R1 reasoning model sent Nvidia stock tumbling and triggered a $1 trillion tech selloff. The Hangzhou startup says V4 can digest one million words of context while matching the performance of OpenAI's GPT-4, Google's Gemini, and Anthropic's Claude. TechCrunch argues the model "closes the gap" through architectural tweaks that make it more efficient than last year's V3.2.

But Bloomberg's first independent benchmarks tell a different story. After running V4 through standard reasoning, coding, and math evaluations, their testers found it still lags GPT-4.5 and Gemini 2.0 by a "clear margin." The gap is smaller than last year, yet the US lead hasn't vanished. DeepSeek's pricing playbook hasn't changed either: API access runs a fraction of OpenAI's rates, continuing the cost disruption that made the startup notorious.

Here's where it gets spicy. DeepSeek tuned V4 specifically for Huawei's Ascend 910B processors, according to Fortune. That matters because US export controls have blocked Chinese companies from Nvidia's latest GPUs since 2022. The Verge notes this pokes holes in the assumption that Chinese AI shops need American hardware to compete.

MIT Technology Review adds a new wrinkle. Their teardown shows DeepSeek sanded down memory usage and bolted on custom low-precision math routines that let Ascend chips punch above their weight class. In plain English: they taught domestic silicon tricks normally reserved for Nvidia's top shelf. The result isn't just a lab demo—it's a signal that Huawei's foundry roadmap can keep scaling if engineers get creative enough.

CSIS analyst Gregory Allen points out the elephant in the room: if DeepSeek somehow obtained banned H100 chips, they'd be admitting to illegal activity. Instead, they've managed to squeeze near-frontier performance from domestic silicon. Reuters reports the model is "adapted to run on" Chinese chips, confirming the workarounds took serious engineering sweat.

DeepSeek's pricing strategy hasn't changed: undercut everyone. Fortune notes V4 continues the company's pattern of offering capabilities at "rock-bottom" rates. Bloomberg's testing confirms the model is "good enough" for most enterprise tasks, which may be all it needs to keep winning business on cost.

The bigger picture? Every dollar OpenAI and Google spend on newer, pricier chips looks a little less necessary if Huawei silicon can keep inching forward. MIT calls this "efficiency arbitrage"—when a cheaper, slightly weaker model wins by being dramatically less expensive to serve. That's the playbook DeepSeek keeps running, and so far, markets keep rewarding it.