The Day in AI

• Canva released Canva Craft on 22 October, its first in-house image generation model, trained on its own design templates and open to all users.
• Moonshot AI released Kimi K3.1 Pro on its API on 23 October, and Artificial Analysis scores it 55, fourth overall on its Intelligence Index.
• Block announced Dessa on 23 October, a speech-to-speech model trained only on human conversation that runs locally on a MacBook Pro M4 Max or better.
• NVIDIA released Nemotron 3.5 Bolt E4B 4GB, an open small model for local voice agents that NVIDIA says needs about 4GB of video memory.
• OpenAI says it built an "offline world" from cached web content so it can train frontier models to browse without live internet access.

Craft works in two stages. It first plans the image (object layout, relative scale, visual hierarchy and any text), then generates it at up to 2K resolution in 30 aspect ratios. Canva trained it on human-made design templates from its own library, such as posters, presentations, infographics and menus. Users can edit one element without changing the rest of the image. Canva has published no benchmark comparison with Google's or OpenAI's image models.

Moonshot AI put Kimi K3.1 Pro on its API on 23 October, the first model in the K3.1 series and the follow-up to Kimi K3. Moonshot says it's a mixture-of-experts model with 2.8 trillion total parameters, 104 billion of them active per token, and a context window of 1 million tokens. Artificial Analysis scores it 55 on its Intelligence Index, 11 points above Kimi K3 (max) and three below Claude Opus 5.5. Moonshot reports 65.7 on Terminal-Bench 4.0, close to its figure of 66.4 for Claude Opus 5.5.

The price is $2 per million input tokens and $10 per million output tokens, and a Highspeed version runs two to three times faster at twice the price. Moonshot says it removed cyber-related data from training to reduce misuse, and it has promised the full weights by 4 November. Until then the model is API only, and nobody outside Moonshot can test the effect of the removed cyber data on the weights.

Block, Jack Dorsey's company, announced Dessa on 23 October. Block says Dessa takes speech in and produces speech out directly, was trained from scratch, and learned everything it knows from human conversation. It runs locally on a MacBook Pro M4 Max or better. A small group of AI safety researchers gets early access first, and Block hasn't said when wider access will follow.

• NVIDIA says it post-trained Nemotron 3.5 Bolt E4B 4GB from Google's Gemma 4 E4B with reinforcement learning only, after supervised fine-tuning hurt conversational quality.
• Amap and Alibaba's ATH group released Lambda-0, a whole-body humanoid model trained on 500 hours of human first-person video, and report 62.5% zero-shot success on a Unitree G1. The code is on GitHub under Apache 2.0, and the source does not say whether the weights are released.
• iFLYTEK launched Agent OS for enterprise agents and the XFYUN CLI, which it says calls more than 700 AI capabilities with one command.
• OpenAI reportedly no longer trains its frontier models on the live internet and uses a cached copy of the web instead.
• Replit moved most of its AI spending from Anthropic to OpenAI in a deal worth more than $100 million a year, according to The Information.
• The UK government reportedly committed $800 million to AI security and biosecurity, including a £100 million biosecurity lab.
• Amazon Web Services capacity for 2027 is "largely sold out", Andy Jassy said, according to The Information.