The Other AI Releases You Missed This Week

The flagship 30B-A3B variant uses a mixture-of-experts architecture that activates only 3 billion parameters at inference, keeping compute costs low while (according to Alibaba’s own RynnBrain-Bench evaluation suite) outperforming Google’s Gemini Robotics-ER 1.5 and Nvidia’s Cosmos-Reason2 across 16 benchmarks. Independent verification of those claims has not yet been published. All seven model variants are available on GitHub and Hugging Face under open-source licences. Google and Nvidia charge for comparable capabilities through their cloud platforms.

ByteDance’s video model generated a voice from a photograph

ByteDance released Seedance 2.0 around 12 February, a text-to-video model that generates realistic footage from written prompts. The model shipped with a feature called Face-to-Voice that ByteDance had to suspend before it even launched properly.

On 10 February, Chinese tech reviewer Tim Pan demonstrated that Face-to-Voice could reconstruct his specific voice and speaking style from a single photograph. No audio sample, no consent, no text prompts describing how he speaks. It inferred vocal characteristics from his face alone. Pan described the experience as ‘terror-inducing’. ByteDance disabled the feature and added a live verification step, but has not disclosed what training data enabled the capability or whether it will retrain the model to remove it.

Separately, a viral video showing a fabricated fight between Tom Cruise and Brad Pitt drew a formal response from the Motion Picture Association, which called it ‘massive infringement’. ByteDance faces simultaneous pressure from privacy advocates over voice cloning and from Hollywood over likeness rights.

Kuaishou upgraded Kling and its share price followed

Kuaishou’s Kling 3.0 extended video output to 15 seconds with improved consistency and added native audio generation across multiple languages, dialects and accents. Kuaishou’s share price has risen more than 50 per cent over the past year, with Kling cited by analysts as a primary driver. The multilingual audio opens export potential in Southeast Asian markets where Kuaishou’s short-video platform already operates.

China’s AI development pace is not slowing down

Three production models from three companies in five days. Alibaba gave its away for free, undercutting Google and Nvidia’s paid offerings. ByteDance shipped so fast it had to pull a feature on day one. Kuaishou turned a model upgrade into a stock catalyst. DeepSeek set this pace in late 2025 when its open-source reasoning model compressed margins across the LLM market, and the rest of the Chinese tech sector has been racing to keep up since.