AI Prompt
BlogAbout
← Back

Qwen4-Max, OpenAI's Lattice Standard Model, and More

4 min read

The biggest release of the weekend landed Sunday evening US time: Alibaba shipped Qwen4-Max, the new flagship of the Qwen family. It's a sparse mixture-of-experts model with 2.4 trillion total parameters and 71 billion active per token, native vision, and a one-million-token context window. Alibaba says the new Qwen4 architecture upgrades the MoE design across attention, residual connections, and embeddings, and that on its internal Work360 benchmark the model beats Qwen3.8-Max while using 35% fewer tokens at high reasoning effort and cutting the average cost per task by around 60% (Alibaba Cloud).

The part developers will notice first: the weights are closed this time. API pricing is $1.50 per million input tokens and $4.50 per million output, and Alibaba's own charts are more mixed than the headline. Qwen4-Max scores 75.7 on DeepSWE v1.1, just ahead of Claude Opus 5.5 at 74.2, but 48.5 on Terminal-Bench 4.0, behind Opus 5.5 at 59.6 (The Next Web). It landed at #6 in Code Arena's WebDev leaderboard (Qwen on X) and is already on OpenRouter (Qwen on X). Alibaba says smaller open-weight Qwen4 models are coming soon (Qwen on X), and Unsloth says Qwen4-27B will run on 16GB RAM setups (Unsloth on X). That smaller release is the one I'm actually waiting for.

The other story I kept coming back to is from OpenAI, which launched a science blog on Sunday. The first guest post is from Nathaniel Craig, who leads the particle theory group at UC Santa Barbara and is a visiting scientist at OpenAI. He describes working with GPT-6 Astra on a candidate nonperturbative definition of the Standard Model on a lattice, a foundational problem that has been open for more than fifty years. Craig says he had a strategy but couldn't prove it worked in every case it needed to cover. Astra turned the strategy into a construction and took the lead on the detailed proof, which went into an October preprint (OpenAI).


What Else Happened

Treasury Secretary Scott Bessent wants answers from Nvidia within 10 days. On Fox News Sunday he asked whether Nvidia was "pressing these seemingly illicit sales to goose their revenues and their earnings," said some of its customers were obviously shell companies for Chinese interests, and urged the board to hire its next CEO from outside the company. (Axios, CNBC).

The Trump administration plans to name Nvidia an AI "national champion." The WSJ reports Nvidia will be one of more than two dozen companies given the designation, which is meant to help US AI firms win business overseas. It came out the same day as Bessent's comments, which is quite the split screen (WSJ Tech on X).

GitHub shipped a security extension for the GitHub Copilot app. You can ask Copilot to find, explain, fix, or dismiss code scanning alerts, and review local changes for new vulnerabilities before opening a pull request. It's in public preview and installs as the github-security plugin (GitHub Docs).

Guillermo Rauch released xeet, an X post renderer written in safe Rust and WebAssembly. It reimplements Blink and HarfBuzz text layout instead of booting a headless browser, and renders a text post to PNG in a few milliseconds. The interesting part is how it's built: an oracle runs Chrome to verify text layout rules, and a CI canary files PRs for agents when the original changes (Guillermo Rauch on X, xeet).

Apple's smart glasses are reportedly behind on software. Mark Gurman says Apple wants to unveil them at WWDC 2027 and ship before the end of that year, but the software is running behind, so both dates are uncertain even though most of the hardware is finished. The new Siri AI integrations are the centerpiece of the push (Mark Gurman on X, Bloomberg).


What I'm Reading

That's it for today. 👋