Three usable things: multimodal inference that fits on the device, OCR small enough to make the hosted premium awkward, and a scoreboard that prints the hardware.
Three items this week share one quality: each can be checked by looking rather than by building. One of them runs in a browser tab. The other two are published numbers with their conditions printed beside them, which is rarer than it sounds. The rest of the week's material sat with people building agent harnesses and compression pipelines, so this is a short edition.
Liquid AI's 2B stack captions video in-browser (procurement).
Liquid AI released four models at once under one architecture family: text at 1.7B, vision at 2B, audio at 2B, and a combined version that takes all three. Every one of them is small enough to sit on a phone, and the compressed builds that consumer tooling actually reads shipped the same day rather than arriving six weeks later from volunteers. A demo already has the vision model captioning live video inside a browser tab with no server call, which is the part worth watching before reading a single benchmark line. Ten days ago the in-browser milestone was text generation. The question this opens is narrow and answerable: for anyone carrying a per-image or per-minute line item for description, captioning, or transcription, whether that capability still has to leave the building now has two credible answers.
Small OCR models close on hosted parsers (procurement).
A 1.3B model scored 93.68 on real-world document parsing, and Baidu's Unlimited OCR 2 arrived within hours of it. The accuracy number is the less interesting half. The release also includes a second model whose job is to classify how a parse failed, by category, instead of reporting one distance score, which turns document quality from something a person spot-checks into something a pipeline can gate on. Two credible small document parsers in a day puts pressure on a specific line item: hosted OCR priced per page. The question that gets sharper on the next renewal call is which documents the vendor's accuracy figure was measured against, because that answer is now comparable to a number someone published for free.
A benchmark table that prints the hardware (evaluation).
Someone published full benchmark suites, not subsets, for compressed open models from 4B to 35B, each score sitting next to the graphics card it ran on and the throughput it produced. Vendor leaderboards copy cloud scores, so they describe a machine nobody in this conversation owns. The useful column here is the hardware one: a score beside its memory requirement and its tokens per second is a cost per token, readable rather than inferred. That matters more this month than last, given where consumer card prices have gone. For anyone weighing a hosted per-token contract against buying the machine, this is a table to read, not software to stand up, which is most of why it belongs here.
Of the three, the browser demo is the one that answers its own question in about two minutes, and it answers it for the exact capability most teams are still paying per call for. The other two are reference material for the next time someone quotes a score without saying what it ran on.