- Tencent's Hy4 arrived with image understanding and computer use, API pricing at $0.834 per million input tokens, and a promise of downloadable weights within days.
- Revolut says it cut its AI coding bill roughly 90% by moving about 2,000 engineers off Claude and Codex onto self-hosted open models.
- A study of 21 million papers found open-weight model use in single-model research hit 44.0% in 2026, but the growth came from Chinese families while Llama and Mistral slipped.
- The New York Times and three Tribune papers sued Brave over its search index, going after the supplier of AI training data rather than the labs.
- Ohio voters approved data center bans or limits in at least 14 towns, including the site of a planned Meta campus.
- If you want the frame underneath all of this, start with where generative AI turns into functional AI.
Today's news all sat one layer below the chatbot. The interesting fights were about who supplies the weights, who supplies the training data, and who supplies the electricity. Model quality barely moved; the plumbing moved a lot. And in three separate stories, the answer to "who owns the pipe" turned out to be someone other than the company whose logo is on the product.
The Front Page: Tencent's Hy4 aims at the download button
Tencent released Hy4 on November 4, the first model in its family that can read images and operate a computer, meaning it can click through a graphical interface or work in a command line rather than only emitting text. It is live on Tencent Cloud and OpenRouter at $0.834 per million input tokens and $2.501 per million output tokens, and the company says the open weights and a system card will follow in the coming days. The Hy team made its pitch in its own research post, and Tencent published a press release claiming a spot among the world's leading open-source models. TechNode reports the model has 770 billion parameters, a figure Tencent did not publish itself.
Hold the scoreboard talk. Nearly every number Tencent offered comes from Tencent. The headline result is a blind evaluation it ran with 213 finance professionals across 579 tasks, where Hy4 scored 3.04 against DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro at 2.68. That is a one-point gap on a three-point-ish scale, graded on a test the vendor designed. The gamedev gains, 19.1% on its GameDev benchmark and 9.8% on GameLogic, are measured against Hy4 preview from August, so they tell you Tencent improved on Tencent. To its credit, the research post concedes that "a notable gap remains" with top proprietary models on research-level math, long-horizon coding and hard visual questions. The weights are the part that matters, and the weights are not out yet.
What it means: The open-weight frontier is being set in China, and the reason is less ideological than logistical. Access to Western compute keeps narrowing for Chinese labs, so shipping weights is how they win distribution they cannot buy. For a builder outside China the practical question is not whether Hy4 beats anything on a leaderboard, it is whether a 770-billion-parameter file you can host yourself changes your cost model. Until the download link appears, treat this as an API launch with a press release attached.
Releases & Features
Revolut moved 2,000 engineers to self-hosted open models. The London fintech says its AI coding costs fell about 90% after it shifted roughly 2,000 developers off Claude and Codex onto Z.ai's open-weight GLM models running on 56 Nvidia B200 GPUs, at around $0.30 per million tokens. David Abitbol, who runs AI operations there, laid out the migration at the AI Engineer conference in October; The Register wrote it up on November 3, and the talk slides show time to first token dropping from 14.26 seconds mid-migration to 1.68 seconds on the same hardware while demand grew 76%. The gains came from tuning, not more GPUs: FP4 quantization, splitting prefill from decode, and capping context around 256,000 tokens. Abitbol's verdict on million-token windows was blunt: "Long context is a scam."
Mistral Large 4 weights are out. The French lab published downloadable weights for its roughly trillion-parameter model on October 29 under a modified MIT-style license, with a separate commercial license required if you resell the model as a hosted service. Details here.
What it means: Revolut's reported 60% improvement in coding time and 70% in lead time are internal figures that do not separate the model from everything else the team changed, so read them as direction rather than measurement. The durable finding is the one about behavior: metered billing made engineers hoard tokens for the end of the quarter, and a flat infrastructure bill let them stop rationing. That is an incentive story, not a benchmark story, and it is the clearest argument yet that per-token pricing quietly shapes how people work. If your own AI spend is unpredictable, the first move is not migration, it is measuring what the spend buys you per task, task by task, before you move anything.
In the Lab
A paper from Zackary Okun Dunivin at the University of Stuttgart asks a simple question with an uncomfortable answer: have scientists really adopted open models? Working from 21 million full-text articles through June 2026, he finds that open-weight use in single-model papers climbed to 44.0% in 2026, and 87.2% of multi-model papers include at least one open family. But the growth is not a turn toward open science. It is a turn toward Chinese models. Use of Llama and Mistral declined from 2025 to 2026, while researchers at Chinese institutions had 2.23 times the odds of picking an open model and accounted for 44.0% of the increase since 2023 while writing 24.8% of the sample. Adjusted use of Chinese open models was 37.1% among papers with Chinese affiliations against 9.2% among papers with no observed China link. Read the paper.
What it means: "Open" has been doing double duty as a technical property and a virtue signal, and this pulls the two apart. Researchers are not choosing inspectable weights because reproducibility demands it; they are choosing whatever is good, cheap and available in their market. That makes the scientific record quietly dependent on which labs happen to be giving models away this year, which is a commercial and geopolitical variable rather than a scientific one. If you cite a result produced with a model that gets deprecated or geo-blocked, your methods section becomes unreproducible for reasons that have nothing to do with your method.
The Oversight Desk
The New York Times sued Brave Software in Manhattan federal court on November 2, and the New York Daily News, Chicago Tribune and Orlando Sentinel filed a closely matching complaint in the same court. The claim is that Brave scraped news articles into a search index of "over 30 billion webpages" without disclosing the name of its crawler, then sold access to that index to AI companies including Anthropic, Perplexity, Mistral, xAI and Cohere. None of those companies is a defendant. The 72-page complaint calls Brave "the supplier, dealer, and distributor of stolen copyrighted content," cites a Brave blog post saying Anthropic used the Search API while training Claude 2, and estimates Brave has earned more than $100 million from publishers' work. It also targets the browser's paywall blocker as a DMCA circumvention tool, which Ars Technica tested and found enabled by default. Brave had not responded publicly as of November 3, and none of the claims has been tested in court.
What it means: This is the copyright fight moving one link up the chain. Suing a model maker over training data is slow and factually murky; suing the search API that feeds it is concrete, because there is an invoice. Lawyer Steven Lieberman put the theory plainly: "Buying stolen goods secondhand doesn't make them legitimate." If that holds, every product that grounds its answers in live web search inherits a licensing bill it has not been paying, and the cheap retrieval layer most AI apps quietly depend on gets more expensive. Worth knowing where your retrieval data comes from before someone asks you in a deposition.
Three stories today came down to a bill nobody had itemized. Before you switch models or hosts, find out what your current AI spend buys you per task. Tell BYOBot the job and get a spec back.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Ohio towns voted the buildout down. Voters approved data center bans or restrictions in at least 14 municipalities on November 3, including Piqua, where a charter amendment capping campuses at 25 megawatts passed 66.7% to 33.3% at the site of a planned Meta project. Fortune.
- A planned chip deal fell through. Oracle's arrangement to supply Tencent was scrapped after US pressure, the Financial Times reported, which is useful context for why Chinese labs keep shipping weights. Summary.
- Steam labels are becoming normal. AI disclosure labels now appear on roughly half of new games on the storefront, a rare case of a platform-level transparency rule sticking. Source.
- Fal raised at a $16 billion valuation. The generative media inference company has now pulled in $1.65 billion across two rounds, a bet that serving other people's models is the better business. Source.
- Google added $15 billion to a Malaysian data center plan for 2027 and 2028, while US towns were busy voting campuses out. Source.
The Bottom Line
Nobody shipped a smarter model today. What shipped was leverage: Tencent betting that a downloadable file beats a better API, Revolut proving the bet pays at scale, a Stuttgart researcher showing science has already quietly taken the same deal, and the Times arguing that the data underneath all of it was never free in the first place. The thing to watch is whether Hy4's weights land this week, because that is the moment the open-weight frontier stops being a talking point and becomes a line item. Either way, the useful move is not picking a side in the open-versus-closed argument. It is knowing what one completed task costs you today, so that when the pipe changes you can tell whether anything got better.
Frequently Asked Questions
-
Open weights means the lab publishes the trained numbers inside the model, so you can download the file and run it on hardware you control. It does not mean free. You still pay for the GPUs, the serving stack and the people who tune it, and the license may forbid reselling the model as a hosted service. Open weights get cheaper as volume rises and stay expensive when it doesn't, which is why a company with 2,000 engineers hammering a model all day sees a very different bill than a team of five. Our rundown of hosted agent runtimes covers the middle ground between the two.
-
Usually not yet. Revolut's reported 90% saving arrived with 56 B200 GPUs, a dedicated AI operations team and months of tuning, which is a different problem from a startup paying a few hundred dollars a month in API fees. The transferable lesson is the measurement, not the migration. Work out what each completed task costs you now, then revisit hosting when the number stops being trivial. Mapping the job first helps, which is what the workflow directory is for.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
