DeAI Daily Brief — 23 October 2026

Today in DeAI: DeepSeek closes a reported $15 billion round, Kyutai ships real-time world models fully open, and China's draft export rules leave open weights exportable.

DeepSeek raises about $15 billion from Tencent and CATL, Bloomberg reports

Bloomberg, citing people familiar with the matter, reports DeepSeek raised about 100 billion yuan ($15 billion) in a round that started with a 50 billion yuan target and doubled along the way, with battery maker CATL and Tencent among the largest investors. The company has not confirmed the round, and Reuters reports details could still change. Why it matters: the capital is earmarked for data centers and open-source models ahead of a possible IPO, which is a direct bet on the open-weights release cadence builders depend on. (Bloomberg) — our full coverage

Kyutai releases CADRE: real-time world models, fully open

The French lab released a family of causal-diffusion video autoencoders for interactive world models — paper, training code, weights, and a playable demo — with world models that react to user actions in under 230 ms at 20 fps on a single H100, per Kyutai's own measurements. Why it matters: the artifact set (code plus weights plus demo) is a worked example of openness beyond a weights-only drop, and the latency findings matter to anyone building interactive video on single GPUs. (Kyutai) — our full coverage

China's draft export rules cover models and compute — but exempt open weights

Reuters reported October 16 that China's draft export-control catalogue covers 23 AI technologies including large models, agents, and compute clusters, while proposing simplified licensing for already-public open-source technology. Why it matters: today's open-weight pipeline from DeepSeek, Alibaba, and peers stays exportable under current drafts — but the exemption's fate is the supply-side trigger to watch. (Reuters)

NVIDIA's Nemotron 3.5 open weights keep coming at a 4-6 week cadence

NVIDIA released the Nemotron 3.5 Super VL 120B (claimed 1M context) and Bolt 74B, sized for a 24 GB GPU, in mid-October — with Bolt positioned within single-consumer-GPU reach. Benchmark numbers are vendor claims or single third-party measurements pending reproduction. Why it matters: a steady cadence of capable open weights in small envelopes keeps shifting the self-hosting math. (NVIDIA)

A Hugging Face listing points to Mistral Large 4 open weights on October 31

Pricing tables already show API rates of $0.68/$2.09 per million tokens "since Oct '26." Why it matters: if the weights land as listed, Mistral repeats its open-weights-with-live-API pattern — and the release date is a listing, not a confirmed announcement, so treat it as a watch item. (Airtics)

Watching tomorrow

The DeepSeek round is set to close "in the coming days" — a company statement or investor filing confirming the 100 billion yuan figure would turn the week's biggest reported claim into fact.