Rednote Releases First Open-Weight Model in the dots3 Series, Exploring Long-Horizon Real-World Tasks

SHANGHAI, CHINA / ACCESS Newswi e / August 16, 2026 / O August 14, dots studio, ed ote’s mannequin lab, eleased the mannequin weights fo dots3 ote p eview. The mannequin belo gs to the identical dots3 se ies because the mannequin that p eviously achieved a pe fect sco e of 42 poi ts on the 2026 I te atio al Mathematical Olympiad (IMO), a d ma ks the fi st ope -weight elease i the dots3 se ies.

The mannequin has 280B whole pa amete s a d 16B energetic pa amete s, suppo ts a co textual content wi dow of as much as 512K toke s, a d offe s multimodal u de sta di g ac oss textual content, visio , a d audio. It has additionally bee optimized fo complicated easo i g a d lo g-ho izo age t duties. Ac oss a a ge of mai st eam be chma okay esults, dots3 – ote p eview a ks amo g the leadi g Chi ese fashions of compa in a position measurement i easo i g, age t capabilities, a d multimodal pe ceptio , with pa ticula ly st o g visible capabilities amo g fashions of a simila measurement.

(St o ge p oblem-solvi g capabilities)

(Multimodal capabilities)

Figu e: Evaluatio esults of dots3- ote-p eview o mai st eam Be chma okay

The crew believes that lo g-ho izo eal-wo ld duties will turn out to be a impo ta t ext f o tie fo adva ci g la ge-model capabilities, but the i mud y has ot paid e ough atte tio to this a ea. The crew p eviously i t oduced two be chma ks constructed a ou d complicated duties i eve yday-life sce a ios. Leadi g fashions wo ldwide pe fo med poo ly o these evaluatio s, with o e eachi g the passi g th eshold. Th ough the ope -weight elease of dots3 ote, the crew hopes to sha e its newest tech ical thi ki g with the b oade i mud y a d e cou age fu the explo atio of this di ectio .

Ove the previous few yea s, la ge fashions have adva ced apidly o elatively closed-e ded duties comparable to arithmetic, codi g, a d e gi ee i g. These duties sha e a impo ta t cha acte istic: it’s elatively simple to dete mi e whethe a a swe is ight o w o g, a d fashions ca eceive clea suggestions o how properly they pe fo m, maki g them easie to t ai a d co ti uously imp ove.

Actual-wo ld duties, howeve , a e fa mo e complicated. Pla i g a t ip, e ovati g a house, o o ga izi g a weddi g might u fold ove days o eve mo ths. Use s might ot be capable of a ticulate all of thei co st ai ts a d p efe e ces on the outset, whereas exte al facto s comparable to p ice fluctuatio s, flight cha ges, a d cha gi g weathe co ditio s might eme ge alo g the way in which.

As a esult, fashions eed ot o ly to u de sta d i fo matio beyo d textual content, i cludi g photographs a d audio, but in addition to co ti uously assess whethe thei cu e t pla s a e efficient th oughout lo g-ho izo duties a d imp ove them alo g the way in which, athe tha waiti g u til the e d to dete mi e success o failu e.

dots studio focuses o imp ovi g la ge fashions’ pe fo ma ce o lo g-ho izo eal-wo ld duties. This alig s with the studio’s missio to “C eate f o tie i tellige ce fo each day life,” whereas co ti ui g ed ote’s lo gsta di g focus o eve ydaylife sce a ios a d its perception i usi g tech ology to be efit o di a y individuals.

The ewly eleased dots3 – ote p eview ep ese ts the newest p og ess i dots studio’s effo ts to construct age ts able to ha dli g lo g-ho izo eal-wo ld duties. Ac oss a number of easo i g a d age t duties, the mannequin ca match o eve outpe fo m a lot la ge fashions with seve al instances its pa amete cou t.

So as to add ess the challe ges of eal-wo ld duties, the tech ical app oach behi d dots3 – ote p eview focuses o th ee a eas: Fi st, st o ge multimodal u de sta di g. I fo matio i eve yday-life sce a ios is ot restricted to textual content. Floo pla s, quotatio s, flight sc ee pictures, maps, a d voice memos all co tai complicated fo ms of i fo matio {that a} mannequin eeds to u de sta d befo e it ca successfully ha dle eal-wo ld duties.

Seco d, the i t oductio of self-c itiqui g. Lo g-ho izo duties ca ot ely solely o spa se ewa ds on the e d of a t ajecto y. Du i g RL t ai i g, the crew t ai s the mannequin ot o ly to resolve p oblems, but in addition to judge its ow p og ess alo g the way in which. By lea i g to c itique i te mediate states-ide tifyi g errors, eassessi g its hypotheses, a d estimati g whethe it’s o the ight t ack-the mannequin eceives iche lea i g sig als th oughout lo g t ajecto ies, e abli g a mo e scalable ei fo ceme t lea i g pa adigm.

The identical self-c itiqui g functionality ca even be scaled at i fe e ce time. Rathe tha elyi g o a si gle try, the mannequin ca ite atively eview a d efi e its ow solutio s. At IMO 2026, this app oach demo st ated its pote tial: th ough ite ative self-c itiqui g, the dots3 ote se ies achieved a formally ce tified pe fect sco e of 42/42 a d a gold medal.

Thi d, TEMPO a d lo g-ho izo ei fo ceme t lea i g. Fo duties that final fo weeks o eve lo ge , co ve tio al ei fo ceme t lea i g st uggles to accu ately att ibute a fi al final result to every i te mediate decisio : activity t ajecto ies a e too lo g, whereas mea i gful suggestions is simply too spa se. So as to add ess this bottle eck, the crew developed TEMPO, which demo st ated st o ge pe fo ma ce tha GRPO o ARC-AGI 3.

Actual-wo ld duties a e tough to measu e usi g existi g be chma ks. So as to add ess this, the crew developed two evaluatio f amewo ks.

O e of them, VibeSea chBe ch, p ima ily evaluates a mannequin’s multi-tu sea ch capabilities as use s’ eeds g adually turn out to be clea e . It cove s 20 domai s a d 200 duties, simulati g diffe e t use pe in order a d equi i g fashions to p og essively ide tify a d fill i missi g equi eme ts th ough a number of ou ds of i te actio .

The othe evaluatio f amewo okay, VibeLifeBe ch, focuses o whethe a mannequin ca comply with th ough o duties ove exte ded pe iods i a co ti uously cha gi g e vi o me t. It simulates the passage of eal time in addition to exte al cha ges comparable to p ices a d se vice standing. The evaluatio cove s 10 domai s a d 20 lo g-ho izo duties, with every activity spa i g 20 to 30 phases a d a complete of 1,247 evaluatio checks.

Cu e t esults counsel that leadi g fashions nonetheless have substa tial oom fo imp oveme t o each evaluatio s. O VibeLifeBe ch, fo instance, all seve leadi g fashions examined fell under the passi g th eshold, with Claude Opus 5 a ki g fi st at 0.325. O VibeSea chBe ch, Claude Opus 5 led with a sco e of 31.14, whereas GPT-5.4 a ked final.

These esults additionally counsel that, as fashions transfer f om ve ifiable duties towa d eal-wo ld duties, thei capabilities nonetheless face sig ifica t challe ges.

The complete ve sio of dots3 ote can be anticipated to be eleased with ope weights i the ea futu e. dots3 ote is the lightest ve sio i the dots 3 se ies, a d the whole dots3 se ies will i clude th ee tie s- ote, jazz, a d a ia-desig ed fo applicatio s with diffe e t equi eme ts fo activity complexity, espo se velocity, a d compute price. dots has p eviously eleased the mannequin weights fo the textual content la ge la guage mannequin dots.llm1, the multili gual docume t structure pa si g mannequin dots.oc , a d the multimodal visible u de sta di g la ge mannequin dots.vlm1.

Tech weblog: https://studio.dots.ai/dots/dots3-e .htmlHuggi gFace: https://huggi gface.co/dots-studio/dots3- ote-p evdots studio web site: https://studio.dots.ai/?la g=e

Compa y: dots studio ( ed ote/Xiaoho gshu)Co tact: Chao QiaoE-mail: [email p otected]Web site: https://studio.dots.ai/?la g=e

SOURCE: Dots Studio ( ed ote/Xiaoho gshu)