The week the weights came with receipts Placeholder

Every week a few hundred things are announced as AI products. This week, by our count, about forty of them could be used by someone who does not work at the company. That is roughly the usual ratio, and it is why this list exists.

The top of the board went to open weights, which is not unusual. What was unusual was the paperwork. Both of the week’s large releases came with a list of what they were trained on, and in one case that list was detailed enough to argue with.

01Receipts, and a model to go with them

Tern-3 topped the week’s board, scoring high on Real and solidly on Useful. It is a 14B coding model that fits on one consumer card, under a licence lawyers do not need to read twice. Our editors pulled the weights and reran the harness; two of three suites landed within a point of the card.

Readers were less impressed on Novel, and they are right. This is a known recipe done carefully. The comment thread spent forty messages establishing that, and then another twenty explaining how they had already deployed it.

Basalt-70B is the more interesting release and the less finished one. The manifest and deduplication scripts are public and thorough; the weights download, but nobody outside the lab has yet reproduced the headline numbers on hardware they can describe. It sits at Pending, and it will stay there until someone does.

A training manifest is a promise you can check. That alone puts it ahead of most launch posts.

02Decisions, on the record

Harlow County now routes non-emergency 311 calls through a triage model, which is not news. Publishing the audit log, including the calls the model sent to the wrong department, is. It was the week’s highest-scoring item on Notable, and the only one to score above 3.5 on all four dimensions.

Northgate Mutual’s claims-triage model card is thinner, but it includes the number regulators asked for: how often a human overrides the model, and how often an appeal succeeds. We would like to see that figure become boring and universal.

03Vapor of the week

Waypoint’s autonomous site-reliability engineer has a demo video, a waitlist and a pricing page with no prices on it. It drew more comments than anything else this week, mostly from people who carry a pager. Its Notable score is respectable; its Real score is the lowest on the board. Both are deserved.

Halyard-2, a vision model with a benchmark table and a “weights coming soon” banner, joins it in the vapor log. We will happily change the stamp the day the download link works.

04What the graph noticed

The topic model flagged a new cluster this week: public audit logs for government triage, seven items across three jurisdictions, none of which cite each other. Meanwhile voice agents nearly doubled in mentions over eight weeks while the number that shipped rose by a third. Most of the gap is outbound sales callers.

As ever: if you used something this week that worked, submit it with the evidence. If you used something that did not, say so in the thread. Both are useful.