DeepSeek V4 Is Coming — And It's Going Multimodal
The Chinese AI lab that shocked the world in 2025 is back with a model that handles text, images, and video.
Remember when DeepSeek R1 dropped in January 2025 and basically broke the AI internet? Markets panicked. Big Tech scrambled. And suddenly everyone was asking: how is a Chinese lab doing this with a fraction of the compute?
Well, they’re about to do it again.
V4: Text, Images, Video — All in One
DeepSeek V4 is expected this week (first week of March 2026), and it’s going multimodal. We’re not talking just text generation or code completion here. This is a unified model that handles text, images, and video in a single system.
That’s a big deal. Most models so far have been siloed — great at text, okay at images, terrible at video. DeepSeek is coming with everything at once.
Trained on Forbidden Hardware
Here’s the kicker: V4 was trained on Nvidia’s Blackwell chips. You know, the ones the US government banned export to China?
Yeah, DeepSeek reportedly trained on them anyway. Whether they smuggled them in, used cloud clusters, or found some other workaround — the fact that a Chinese model was trained on the most advanced AI hardware available is… consequential.
What Made R1 Scary
If you missed the R1 saga: DeepSeek released R1 in January 2025 and it matched or beat OpenAI and Anthropic on reasoning benchmarks — while being open weights and dramatically cheaper to run. The release “rattled global AI markets” (Reuters’ words).
The stock reaction was immediate and visceral. Nvidia dropped. Big Tech went into crisis mode.
V4 is the follow-up. And if the pattern holds, it’s going to matter a lot.
Why This Matters for You
If you run local AI: DeepSeek models have been among the best performers in LM Studio and other local runners. V4 being multimodal could mean running image and video understanding locally — on your own hardware.
If you just watch the industry: DeepSeek keeps proving that the “compute = capability” assumption might be overblown. Smaller, smarter, cheaper keeps winning.
Sources
- DeepSeek V4 Multimodal Launch Imminent — Evermx
- DeepSeek Wikipedia — Wikipedia