AI Daily Brief — 24 April 2026

A double-headline Friday. DeepSeek drops V4-Pro — a 1.6-trillion-parameter Mixture-of-Experts model (49B active) with a 1M-token context window — and a lighter V4-Flash (284B). Both ship under MIT, both support 1M context. Google Cloud Next ’26 closes with the Virgo Network scale-out fabric and a full agentic recap.

Top stories

• DeepSeek-V4-Pro launches open-source. 1.6T total / 49B active params, MIT licence, 1M-token context. Designed for advanced reasoning, complex SWE, and long-running agentic tasks. In the 1M-token context setting, V4-Pro reportedly needs only 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2. via DeepSeek API Docs
• DeepSeek-V4-Flash. The lighter 284B-parameter sibling, also at 1M-token context, built for faster, lower-cost inference. via Hugging Face
• Google Virgo Network. Google’s scale-out AI data-centre fabric that underpins the AI Hypercomputer — the rails behind the TPU 8t/8i announcements earlier in the week. via Google

By the numbers

• DeepSeek-V4-Pro: 1.6T total · 49B active · 1M context · MIT
• DeepSeek-V4-Flash: 284B total · 1M context · MIT
• V4-Pro vs V3.2 at 1M context: 27% single-token FLOPs · 10% KV cache

Compiled by AI Feed’s editor from publicly reported announcements on 24 April 2026.