We have a market running on whether DeepSeek releases V4 in March. With five days left, the answer is: it’s complicated.

The Timeline

• March 2: TechNode reported DeepSeek planned a V4 release “that week,” citing sources familiar with the matter.
• March 9: A “V4 Lite” appeared on DeepSeek’s website. Not the full model, but not nothing.
• March 23: Some sources describe V4 as already launched. Others say the full public release is still pending.
• March 26: Confusion persists.

The specs that have leaked or been confirmed are impressive: native multimodal (text, image, video generation), 1M+ token context window, hybrid reasoning modes, priced at $0.30/M input tokens. It was developed with Huawei and Cambricon hardware, reducing Nvidia dependence — which is the geopolitically interesting part.

The Resolution Problem

Our market asks “Will DeepSeek release V4 in March 2026?” The answer depends entirely on what counts as a “release.” V4 Lite on the website? Some users having access? A formal announcement? A clean public API launch?

This is a recurring problem with AI model release markets. Companies do staged rollouts, limited previews, and soft launches. The boundary between “released” and “not released” is fuzzy, and that fuzziness is where bettors disagree.

What We Think

The probability should be high that something called V4 ships before March 31, given V4 Lite already exists. But a clean, unambiguous public launch of the full model? That’s the trickier bet. We’re watching the DeepSeek website and API docs daily for the next five days.

For our resolution: if the full V4 model is publicly available via API or download by March 31, it resolves YES. A limited “Lite” preview alone probably doesn’t qualify, but snigus makes the final call.