Today we introduced a novel system one model, Clairic, based on regression instead of classification (i.e. decision models like Clef). This enables estimation of continuous values without having to exhaustively list choices upfront. It also enables very efficient extraction of structured data from unstructured inputs. Open weights coming soon!

Will this be a pretrained foundation model replacement to one of my favorite @GoogleDeepMind papers?

system one models are still just clever autocomplete

How it compares with GliNER 2.5 ?

lets gooo

Super interesting

Gave this a shot similar to ScaleDown last week. Recall too weak for extraction use cases 83% recall vs 94% for luna max, precision was 56% iirc played around with the prompt etc. Heavy extraction use case for

@thenewgenesis04 market prediction, worth a try.

Will you make it available on @OpenRouter ?

@rsuyoy

I really like the estimate type; which you can provide a min / max with and get a median back with an 80% range. Looking forward to the open weights.

Cool stuff. Once weights are out, any plans for LoRA or PEFT support?

oh I like this, “how much” feels like a pretty natural next step from “which one”

First thought lol

@bot please review this thoroughly and then deeply consider how we work. Get with me one-on-one with your thoughts on the value of implementing it and what you would propose. Looking for low-hanging fruit, slam dunk, alpha

Regression over classification is a smart move for open-ended inputs. Curious how it handles calibration — continuous outputs seem harder to trust than a probability distribution over known choices.

are you hiring?

回归框架讲连续值、不用事先穷举选项，这点很关键。想问下 estimate 没证据时区间会不会故意放宽？和 Clef 级联时，你们更常先抽再决策，还是偶尔反过来？

Extracting a big one. Will take a look at it for my projects

A promising shift—regression could make system-one reasoning more flexible and efficient when the possible outcomes aren’t predefined

The shift sits in the loss function: classification discretizes, regression preserves the gradient. A system-one model on regression bets fast intuition is better served by continuous estimation. Question: who scores a regression-based intuition, and against what ground truth?

Continuous values meaning...numerical ranges?

do you have any tips for training on these?

Does it match or beat Jev

Noice, cloudflare building cool stuff. Give mouse-1 a look too

@grok 你给我介绍一下它是什么原理？什么什么作用？然后应用场景

@grok what the hell is regression. explain el5

Regression instead of classification for extraction is the kind of boring-sounding idea that ends up eating everything. I used to burn a giant LLM just to pull invoice dates out of PDFs. Reading this made me realize most of my AI workload was never reasoning at all.