- Alibaba's 2.4-trillion-parameter Qwen3.8-Max goes open-weight next week, alongside a 27B version that ordinary hardware can host.
- Cohere open-sourced a coding agent under Apache 2.0 with 30 billion parameters total and only 3 billion active, small enough for one machine.
- Epoch AI's benchmark of genuinely unsolved math problems grew to 50. AI has cracked three, and none of the hard ones.
- Cursor and Y Combinator both put working agent infrastructure on GitHub with real licenses attached.
- Congress is still deadlocked over whether federal AI rules should switch off state ones, while Europe's high-risk obligations already came due.
- The model worth copying today is the small one. A 3B-active coder you host yourself covers most of what sits inside an agent workflow.
Openness had a great day on paper. Weights got promised, licenses got attached, repositories went public, and four different companies got to call themselves open by lunchtime.
Look at what stayed shut. The flagship weights land in a form that needs a rack of machines to load. The benchmark measuring whether any of this works is a paid product, and so far it has one customer. The rules that would settle who answers for a bad outcome are stuck in committee. Publishing a file is the cheap part of being open. Today was a good reminder of which parts are expensive.
The Front Page: Alibaba's 2.4-Trillion-Parameter Model Goes Open, Sort Of
Alibaba made Qwen3.8-Max broadly available on August 3 and confirmed the open weights ship next week. It's a 2.4-trillion-parameter mixture-of-experts model, meaning it routes each request through a fraction of itself instead of firing the whole thing, and it takes text, image and video in while returning text. The model page lists a one-million-token context window and pricing of $2.00 per million input tokens, $6.00 per million output, dropping to $0.25 on cached input. A second checkpoint, Qwen3.8-27B, is also going open-weight. This would be the first Max-class Qwen released openly, and it marks Alibaba walking back a year of keeping its best models closed.
Now the fine print. At 2.4 trillion total parameters the checkpoint needs a rack of machines, and Alibaba hasn't published how many parameters fire per token, so nobody outside the company can model what it costs to serve. On Alibaba's own benchmark table it scores 86.6 on Terminal-Bench 2.1 and 92.6 on GPQA Diamond, a whisker above its predecessor's 92.4. The multimodal comparisons run against Qwen3.7-Plus rather than Qwen3.7-Max, which makes the generational jump look bigger than it is. Alibaba's own reinforcement learning curve peaks near 4,000 training environments and then slides back down. Credit where it's due for publishing that chart at all.
What it means: "open weights" has quietly become a headline rather than a handover. A file you can download but can't load is a press release with a checksum. The interesting checkpoint here is the 27B, because that one fits on hardware companies already own, and that gap between announcement and access is worth tracking every time a lab uses the word open. When you evaluate a release, skip the parameter count and ask what runs on the machines you have.
Releases & Features
Cohere open-sourced a coding agent that fits on a single GPU. North Mini Code is a mixture-of-experts model with 30 billion parameters total and 3 billion active, released under Apache 2.0, with a 256,000-token context window and weights on Hugging Face. Cohere says it was trained for agentic coding from the start rather than adapted from a general-purpose base, and that it hits up to 2.8 times the output throughput of Devstral Small 2 on matched hardware. That's the company's own number on the company's own setup, so treat it as a starting point. The useful part is the shape: 3 billion active parameters means a repo-level coding agent running on your laptop or one rented card, with your source code never leaving the building.
Underneath the models, the plumbing went public too. Cursor released Mixture-of-Kittens, a deterministic training kernel for mixture-of-experts models on Nvidia GB300 racks. Deterministic means the same inputs produce bit-identical outputs every run, which sounds like a footnote and is really the difference between a training run you can reproduce and one you can only describe.
What it means: the frontier gets the headlines and the small releases get used. A 3B-active model under Apache 2.0 is something a two-person team can put in production this week without asking legal or finance for anything.
In the Lab
Epoch AI expanded FrontierMath: Open Problems to 50 problems on July 31. These aren't exam questions with hidden answer keys. Every problem has to be genuinely unsolved, hard enough that at least two professional mathematicians have already failed at it, interesting enough that a solution would be publishable, and verifiable by a computer program in under an hour. That last requirement is what makes it work: nobody knows the answer, but a machine can check one.
The scoreboard reads three solved, 47 still open. Epoch sorts them by how big a deal a solution would be, and that breakdown is the story. Two of the three cracked problems sit in the "moderately interesting" tier and one is a "solid result." The tiers labeled "major advance" and "breakthrough" are still at zero for nine.
What it means: this is the rare benchmark that can't be gamed by memorizing the training set, which makes it the most honest capability signal available right now. It's also a business. Epoch sells access to the verifiers to fund the work, and it notes plainly that OpenAI is the only entity to have bought that access so far. Independent measurement funded by the companies being measured is the structural problem sitting under every AI leaderboard, and this one at least says so out loud.
The Oversight Desk
America's attempt at a single federal AI law is going nowhere, and the reason is instructive. The Great American AI Act, released as a discussion draft on June 4 by Representatives Jay Obernolte and Lori Trahan, trades new federal safety, transparency and audit duties for the largest model developers against a three-year freeze on state laws that regulate how models get built. Preemption is the word for that freeze: federal rules switching off state ones.
The draft got bipartisan skepticism from both directions at once. Many House Democrats think the preemption is too broad. Many House Republicans think it's too narrow, since it leaves state rules on everything past deployment untouched. With the legislative calendar thinning out, it's unlikely to move as drafted this year. Meanwhile the EU's high-risk obligations under its AI Act came into application this month, so the developers waiting on Washington are already complying with Brussels.
What it means: plan for a patchwork, because that's what you're getting for at least another year. If you operate across states or across the Atlantic, build to the strictest rule you touch and stop hoping a federal standard arrives to simplify it. The documentation you'd need for Europe covers most of what any US state has asked for so far.
Today's lesson is that the biggest model on the leaderboard is rarely the one you should run. Tell BYOBot what the job is and it'll spec an agent around a model that fits your hardware, your budget, and your data rules.
On the Radar
Smaller moves worth a glance, with the sources if you want to go deeper.
- Y Combinator open-sourced QM, an MIT-licensed harness for running several agents together inside Slack or a browser, so a team watches the same agent session instead of everyone prompting alone. MarkTechPost.
- Genspark released GenOffice, a free, ad-free AI office suite for macOS and Windows with docs, sheets, slides and PDF handling. Worth watching how long free lasts. MarkTechPost.
- Reflex open-sourced XY, a Rust-backed Python charting library that keeps 100-million-point plots interactive. Unglamorous, and exactly what analytics agents choke on today. MarkTechPost.
- A judge granted final approval to Anthropic's $1.5 billion copyright settlement, covering roughly 500,000 works at about $3,000 each, the largest US copyright settlement on record. Insurance Journal.
- China's rules for AI companion services are in force, issued by the Cyberspace Administration with four partner agencies and covering any service that simulates personality to provide ongoing emotional interaction. Latham & Watkins.
The Bottom Line
Two models went open today and only one of them changed what anyone can build tomorrow. It wasn't the trillion-parameter one. The pattern holds up across the whole edition: the moves that matter to people doing the work are small, licensed, and sized for hardware that already exists, while the headline numbers keep climbing somewhere you can't follow them.
So read the releases from the bottom of the leaderboard up. Ask what fits on your machines, what license it carries, and what you'd have to prove to a regulator if it went wrong. Then go build one small thing with it, because a 3-billion-active model doing one job well beats a 2.4-trillion one you're only ever going to read about.
Frequently Asked Questions
-
Qwen3.8-Max is Alibaba's largest model to date, a 2.4 trillion parameter mixture-of-experts system that takes text, image and video in and returns text, with a context window listed at one million tokens. Alibaba made it broadly available on August 3 and says the open weights ship the following week, which would make it the first Max-class Qwen released openly. Running it yourself is a different question. At 2.4 trillion total parameters the checkpoint needs multiple machines, and Alibaba hasn't published the activated parameter count, so serving cost can't be modeled yet. The companion checkpoint, Qwen3.8-27B, is the one sized for hardware most teams already have. If you'd rather not host anything, our guide to hosted AI agents walks through the tradeoff.
-
Open-weight means the trained numbers inside the model are published for download, so you can host it yourself instead of calling somebody's API. It isn't the same as open source, because the training data and training code usually stay private. Active parameters are the number that decides whether hosting is realistic. A mixture-of-experts model routes each token through a small slice of itself, so Cohere's North Mini Code carries 30 billion parameters in total but runs only 3 billion at a time, which is why it fits on one machine. Total size tells you what you have to store. Active size tells you what you pay for on every single request. Yesterday's August 4 edition covered the flip side of open weights, which is what happens when anyone can point one at the open internet.
-
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.
