AI briefing: November 4

1Tencent launches Hy4, a new open-source heavyweight

Tencent, the Chinese tech giant behind WeChat, announced Hy4 today, its most capable AI model yet. It is the first in its family that can understand images and charts, and the first with "computer use" - the ability to operate real software interfaces, including professional tools like Blender and CAD, rather than only chatting. Tencent says it will publish the model's downloadable weights in the coming days, which would place it among the strongest "open-source" models anyone can run themselves.

In plain English: an AI model is the trained software behind a chatbot. "Open weights" means its learned internal settings can be downloaded and run on your own computers - getting the engine, not just renting the ride. And computer use means the model can click, type, and complete multi-step work inside ordinary programs.

How good is it? In Tencent's own blind evaluation with 213 finance professionals, Hy4 scored highest among open models on real financial tasks (3.04, ahead of DeepSeek's V4.1 Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68). Tencent also says internal testing cut its hallucination rate from 12.95% to 7.2%. These are the company's own numbers, not independent measurements. Pricing stays at about $0.90 per million input tokens, cheap by frontier standards.

Why it matters: the open-model race - DeepSeek, Alibaba's Qwen, Z.ai's GLM, and now Tencent - keeps closing the gap with the paid frontier models, and open models are what many startups and governments actually run. A stronger free option pressures everyone's prices.

2Fal raises $1.3B to run other companies' AI, doubling its value to $16B

Fal, a San Francisco startup, raised $1.3 billion in a round led by Bessemer Venture Partners, with Nvidia, Sound Ventures, and T. Rowe Price participating - valuing it at $16 billion, double its April valuation, Bloomberg reported. It announced two rounds at once, $1.65 billion combined, and passed $800 million in annualized revenue earlier this year.

In plain English: Fal sells "inference" - the computing work of running an already-trained AI model to produce answers, images, or video. Instead of buying expensive chips, companies rent Fal's servers to run models (often open ones like the Tencent model above) for their own apps.

Why it matters: this is the picks-and-shovels layer of the AI gold rush, and investors are paying up for it - rivals like Baseten, Fireworks AI, and Modal are growing fast too. One caution from the reporting: these businesses' gross margins stay thin because the computing power they resell is so expensive. Big valuations, unproven economics.

3OpenAI says ChatGPT now handles a fifth as many searches as Google

At a press event Monday, OpenAI's consumer product lead Adam Fry said the number of search queries run through ChatGPT now equals 20% of Google's volume. The figure is self-reported, and Fry did not say where the Google comparison number comes from. For context: Microsoft's CEO said in 2023 that Bing held about 3% of the search market after more than a decade of trying.

In plain English: for a growing share of questions, people "ask Chat" instead of "Googling it." Search is the front door of the internet - and the advertising business attached to it is one of the most profitable ever built.

Why it matters: OpenAI is already testing sponsored product images inside ChatGPT answers as it eyes a possible IPO early next year. Fry told reporters the ads "have not negatively impacted the product." If question-answering shifts from search engines to chatbots, the money shifts with it - which is exactly what Google argued in its own antitrust trial.