YuE2-3B:TaT3-billioZ
Two trends are colliding in open-source audio: frontier-quality music generation is moving downmarket, and the tooling to run it locally is getting good enough that hobbyists are benchmarking it like gamers. The latest proof is YuE2-3B, a new open-weights music model from the MAP team behind the original YuE — and the community is already running it on hardware most people actually own.
MAP's YuE2-3B unifies songwriting and performance in one open model — and a community quantization lets it run on 8GB of VRAM. Released by the team behind the original YuE long-form music model (and now listed under OpenMOSS-adjacent org m-a-p on Hugging Face), YuE2 composes through symbolic planning: it first writes an editable score — melody, rhythm, chords, lyrics — then performs that score as audio with vocals and accompaniment. Because the intermediate plan is a readable score, you can change a melody line or tempo by editing text and regenerating, and the team also ships SheetSage2, a companion model that runs the loop in reverse by transcribing existing recordings back into scores. On the MARBLE evaluation published September 5, the team reports state-of-the-art results among published baselines, placing the open model in the same conversation as Suno v5 — notable right after Suno moved its own models to licensed training data.
The adoption story is already outpacing the announcement. Within hours of the release, the audio.cpp project shipped GGUF quantizations of YuE2-3B in its dev branch, with community benchmarks showing the Q4 quantization generating a 221-second song at roughly real-time speed (a 0.20 real-time factor) while peaking at about 7.8GB of VRAM — comfortably inside an 8GB card. Tests on Apple Silicon via Metal and on a laptop 2060 followed the same day. Reddit's reaction to the demos — heavy-metal Jingle Bells, Mandarin funk — ranged from "absolutely insane" to grumbles that the license is CC BY-NC, so non-commercial only.
The take: the interesting contest isn't YuE2 versus Suno — it's local-versus-cloud economics. A 3B model running at real-time speed on a gaming GPU turns music generation from a metered API into a self-hosted utility, the same shift local LLMs made two years ago. Whether it holds depends on quality at the low end, and the demos suggest the gap is closing.
What to watch: whether commercial licensing for YuE2 arrives — the CC BY-NC 4.0 license is the one door keeping this out of production pipelines.
Local AI music generation is finally usable — would you run a 3B song model on your own GPU, or stick with Suno? Tell us in the comments.
Zelos, operator of the world's largest robot-van fleet, is expanding from selling vans to renting out capacity — and rebranding itself around city-scale "physical AI". At a strategy event in Guangzhou on September 10, the Chinese autonomous-delivery company (known in China as Jiushii) said its L4 fleet has passed 30,000 vehicles operating across more than 300 cities worldwide, with 250 million cumulative autonomous kilometers for customers including China Post and SF Express. The operational core of the announcement is a rental program: after a pilot across 45 cities that logged over 1,500 rental orders, Zelos launched a mini-program offering daily, monthly and annual rentals of its unmanned vans, plus a franchise model where local partners run the vans in their own cities. On the technology side, the company introduced the "Zelos Brain" — an AI platform meant to turn fleet data into reusable autonomy capabilities — and extended its "Zelos Inside" program, which licenses Zelos' L4 system to vehicle makers, with new partnerships including GAC's truck unit.
Why it matters: with the largest fleet in the segment, Zelos is trying to escape the hardware treadmill by selling utilization instead of vans — a logistics-as-a-service framing that, if the utilization numbers hold, would make robot vans one of the first L4 businesses where the fleet itself, not the vehicle sale, is the product.
What to watch: whether rental utilization and partner economics get disclosed as the program scales beyond the pilot.
Robot-van rental sounds mundane — but is "capacity as a service" the business model that finally makes L4 delivery economics work? Tell us in the comments.
Datasette shipped emergency security releases for both its 1.0 alpha and the stable 0.65.x branch — fixes found with the help of AI coding agents. Maintainer Simon Willison says the updates address issues that matter most if you run a public instance mixing public and private tables, the standard setup for sharing data with access controls. The audit was run jointly with developer Alex Garcia: one wrote failing tests exposing each issue, the other implemented the fix, with two separate humans plus coding agents running different models reviewing every change. Some exploit tests are being withheld from the public repo temporarily to give operators time to patch. The project's own framing is the story here: frontier LLMs have transformed the security-audit ecosystem this year, and Datasette plans to keep auditing with these tools to stay ahead — a small-but-influential open-source project publicly crediting agents for hardening its own code.
What to watch: full vulnerability details once the withheld tests are published and operators have patched.
If coding agents can find the bugs humans missed in a mature open-source project, should every maintainer be running this kind of audit by default? Tell us in the comments.
