No. 3224 Ɣuct 20265 min
Mačči d Imudam.D Ssaɛa.
Yiwen umudam war isem deg OpenRouter ibeddel tamuɣli-w ɣef wamek yezzerb waya. Mačci d benchmark. D amudam ačini war isem i yewwḍen ɣer umḍiq amezwaru uqbel ma yeẓra ḥedd anwa i t-ixedmen.
Yal yiwen imnaɣ ɣef anwa umudam ifazen aggur-agi. D amnaɣ ur nṣeḥḥin ara.
Ayen i ibeddlen tamuɣli-w ɣef wamek yezzerb waya mačci d benchmark, mačci d tuffɣa n usnas (release). D yiwen umudam war isem deg OpenRouter.
Amudam war isem yeffeɣ-d. Ulac laboratwar, ulac tazrawt, ulac tabrat n usenker. Ula d yiwen ur yessin anwa i t-ixedmen.
Yewweḍ ɣer umḍiq #1 s useqdec-is, yesɛeddi DeepSeek s ugar n snat n tikkal.
Sin isimanen mbaɛd, Z.ai teqqar-d d ayla-s, tessufeɣ-d weights-nni s yisem n GLM-5.3-Flash, yerna teqqar-d, qrib s war-ma tebɣa, belli tuɣ tesselkam-it ɣef iciban ittwaxedmen di Ccinwa seg tazwara.
- Model
- OX ALPHA
- Lab
- UNKNOWN
- Weights
- UNKNOWN
- Chips
- UNKNOWN
- Usage
- #1, 2× DEEPSEEK+
Ɣeṛ-itt tikkelt-nniḍen. Tazdegi teqreb ɣer frontier. Weights ldin. Mačci ɣef Nvidia. Yerna imuhendisen ufan-t yerna sṭutren fell-as s uḍfar-nsen uqbel ma iɛlem-iten ḥedd anwa-t, acḥal i tesswa deg uselmed-is, neɣ anwa ashag i t-yellan fell-as.
D taqsiṭ meṛṛa. Mačci leaderboard. D tidet-nni belli lfaṛeg yella yakan yeṃdel yerna ula d yiwen ur t-yemmuges alamma d asfaylu n yigmaḍ i t-id-yennan.
Akken ad tfehmeḍ ugar, aya ur d-yekki ara seg ulac: GLM 5.2 tessufeɣ-d deg yunyu s weights ldin akked licence MIT yerna teqqim kra n inaḍufen ɣef frontier iɣelqen, yerna GLM 5.3 tedda-d deffir-s deg ɣuct mbla aselmed amenzu amaynut s lekmal. Yal rrbeḥ i d-yekka seg uselmed n mbaɛd. Maca tuffɣa n isnasen mačci d ayen yesɛan azal dagi. Ox Alpha i yesɛan azal, axaṭer d nettat i d-yusan mbla tacreḍt (unlabelled).
Yella ulɣu dagi, yerna yezweṛ ɣef AI.
Ccinwa tekcem ɣer ssuq. Ccinwa terbeḥ deg ssuq-nni. Sakin ayen-nni yuɣal irxes armi yezmer yal yiwen ad t-yaɣ. Ipanlen imṣebḥiyen. Batteries. Tilifuntat. Drones. Takrusin taladbiya. Yal tikkelt, tallɣa-nsen tettuɣal d yiwet.
Taqsiṭ tagarbit ɣef waya teqqar-d belli d amsel. Isubventiyunen, dumping, ṭṭabla ixeddmen s ttlaḥin. Taqsiṭ-nni tegleḍ deg tuget-is, yerna aya seg wakud aya. Ssuma tedda ɣer wadda axaṭer aḥenṭeḍ iɣelli yelha, yerna deg kra n iseggasen ineggura tazdegi tedda yid-s. Mačci d irxis kan tura. Ttmizzilent ɣef ufaris s yiman-is: tazdegi, aẓayer, ssuma akk deg yiwet n tegnit. D win yemxalfen seg umezzugel ɣef d acu i mazal-d mmeslayen medden.
Sinef aya ɣer tmusni yerna ad tafeḍ tagmuḍt tban-d. Aɣawas-nsen n tira n tengalt yesɛa yakan ssuma i yezgan azal n uɛecraden n Claude Code. Wagi mačci d wadda. D amḍiq n ubrid kan.
Ɣur-i ad ifak dinna anda ifukk umeggi ussnidi. Ilelli i useqdec. Mačci d ilelli deg tebni: d ilelli i useqdec. Kra n yiwen yetteṭṭ ssuma axaṭer lweɛi, tuẓẓla neɣ isefka sɛan azal ugar n lfatura n inference. Nwala afilm-agi yakan.
Epoch AI teṭṭafar aya s wudem iwatan, ihi ad sseqdaceɣ iɛdaden-nsen deg umkan n wid-iw.
Lfaṛeg gar ldin d yeɣlaqen yeɣli. Deg 2024, tanezzayt gar umudam ldin ifazen d win yeɣelqen ifazen tella azal n useggas. Ɣer taggara n 2025, azal n kraḍ n waguren. Deg uzgen amenzu n 2026, azal n ukuẓ n waguren. Aseggas yuɣal d rebɛi deg tmanyaɛcṛa n waguren.
Tazmert n frontier tewweḍ ɣer hardware n umsaɣ deg azal n useggas. Akken i tt-id-tessekyed Epoch: imudam i yezmeren ad qqimen deg GPU n umsaɣ ssawḍen tazmert n frontier deg azal n useggas neɣ drus.
Tazmert n usnerni tettnerni s waṭas. Tazmert n tḥemmimt deg uselmed amenzu tettnerni s azal n 3 imuren yal aseggas: tazmert taɣara s uɛecraden n tḥemmimt, tettuɣal ɣer sin tikwal yal sebɛa waguren ntusɛ. Tazmert n uciban i uduular tulin s 49% yal aseggas seg 2023. Yerna tazmert n frontier s timmad-is teǧwiḍ: aɣmis n tezmar yuli s azal n 8 inaḍufen yal aseggas uqbel Yebrir 2024, yerna azal n 15 seg mi d-ttwaqebḍen imudam n uxemmem d usemɣur n RL.
Sdukel-iten meṛṛa. Deg 2024, teḥwaǧeḍ amudam n 70B akken ad tesɛuḍ tazdegi n GPT-4. Deg 2026, Phi-4 yeddu ɣef 8GB n VRAM yerna yemsawa d imudam i t-yifen s kraḍ tikwal deg tmesnaḍt. Alugan n 27B n Qwen yewweḍ ɣer 77% deg SWE-bench s 18GB. Amudam n 3B yeddu s 30 n tokens s tsint ɣef MacBook Air yerna yezmer ad iqqen ɣer iPhone.
Frontier tettnerni s zzerb. Wadda ittali s zzerb ugar.
Kraḍ n iseggasen.
Ayen ara yeḥsben d tamusni n frontier deg wakud-nni (aswir n Opus ass-a, kra i t-yifen mbaɛd) ad yeddu s wudem adigan. Ɣef tlifun deg tqecwalt-ik d laptop-ik ɣef ṭṭabla. Mačci "yelha kan i twuriwin timecṭuḥin". Ayen n tidet, yeddan s 99.9% n wayen i sutren medden i tmacint: tira, tangalt, tasleḍt, aheggi, aselmed, tasuqilt, kullec.
Ḥseb asnerni-nni yal aseggas yerna kraḍ n iseggasen mačci d aɣtas iɛeggḍen. Tazmert n 3 imuren yal aseggas d azal n 27 imuren deg kraḍ n iseggasen. Rnu-as tazmert n iciban yerna tewweḍ ɣer 89 n yimuren deg ssuma n yal aḥric n tazmert, uqbel ma yesnulfa ḥedd kra icebḥen. Ttmesɣur ittḥettmen deg tmacint-agi d takatut, mačci d FLOPs, yerna asezdeg d nettan s timmad-is i d-yettawin abrid akken ad t-yaɛber: akken i 70B n 2024 yuɣal d 8B n 2026.
Same capability for a third of the compute, every year. The binding constraint on a phone is memory, and distillation gets around it: 2024's 70B became 2026's 8B.
Ayen ara yeqqimen deg asigna d aneggaru n 0.1%. Anadi n frontier. Iagentes n webrid ɣezzifen i yeḥwaǧen amelyun n tokens n usatal d wass n usemɣer n usentay (wall-clock). Yal ayen anda lfaṛeg n 5 inaḍufen deg benchmark d ayen i d-yebnan afaris meṛṛa. Ssuq-nni ad yili d win n tidet d ameqran yerna ad yili d ssuq n yimusnawen kan, akken supercomputing i yella d ssuq n yimusnawen. Yal kra niḍen d sselɛa i tesɛiḍ yakan.
Ma yella tamusni tella deg tmacint yerna ssuma n usnas asigni (hosted) yeddu ɣer 0, amudam yeḥbes ad yili d aḥraz. Kraḍ n tɣawsiwin uɣalent d aḥraz deg umkan-is.
Tuẓẓla. Win ara yessellken amudam adigan amagnu ɣef tlifun, laptop d browser, ad yerbeḥ akken yerbeḥ umsekyed amagnu.
Isefka. Yal amudam ad yili yesɛa tamusni tqerb ɣer wayeḍ. Lxilaf d acu i yessen fell-ak, ɣef takebbanit-ik, ɣef imsaɣen-ik, d ma yella tesɛiḍ lḥeqq akken ad ten-tesqedceḍ.
Iɣezdisen. Azal yeddu s ufella deg tsedda ɣer webrid aneggaru: abrid n uxeddim, aseddukel, annar. Akken i yezga yettili mi ara yerxes wayen yellan ddaw-s.
Ma yella amudam yeddu s wudem adigan, akken daɣen i yeddu wayen akk yessen fell-ak.
Isefka-ik ḥebsen ad ilin d ayen ara tesseliḍ yerna bdan ad ilin d ayen ur d-yeffɣen s babat. Tasnigi i yerran amsekyad yesɛa azal, ifuyla-ik, iznan-ik, azemz-ik, tirmit-ik, d nettat s tidet i ɣur-k drus n ssebba akken a tt-tazneḍ ɣer wanda ma yella, mi ara yili umudam yeqqim ɣef yiwen uciban akked isefka.
Yelha i keč. D ugur i tesnaɛt meṛṛa.
Axaṭer azal n isefka-nni ur yettader ara. Amkan n uweḍ ɣur-sen kan i d-yettadren. Ihi ttɛas ad d-yaweḍ usali ɣer webrid amcum ugar. Isteqsiyen n turagt i yettwarra akken ad ten-tqebleḍ mbla ma tɣileḍ. Isnasen ilelli i deg afaris s tidet d asdukel n isefka. SDKs yeffren ddaw kraḍ n twuriwin. Buttons n "ɛawen-aɣ ad nesnerni ameslub-nneɣ" i yellan yermed s lɛadda. Ɣef acḥal ara yeqreḥ uweḍ ɣer isefka, ɣef akka ara yernu unadi fell-asen s tirawin timaynutin, yerna ɣef akka ur d-yeḍru s tafat n wass.
Sakin agmuḍ wis sin, i ɣur-i ur nettwaḥsib ara akken ilaq.
Anadi yeḥbes ad yili d keč.
Ass-a, mi ara tnadiḍ, asteqsi-nni yasef s ubrawzer, cookie, amiḍan, IP, akked profile i d-yettjemmaɛen seg mraw d xemsa iseggasen. Mi ara yili umudam-ik adigan i ixeddmen anadi, ayen i d-yasfen ɣer yidis-nniḍen d tamacint i d-yesteqsayen. Ur teḥwaǧ ara timagit-ik akken ad tesɛu tiririt, yerna ulac-as ssebba akken ad tt-taweḍ. Mraw d sin n yisteqsiyen seg-k uɣalen d xemsa n yisteqsiyen ur d-ibanen d kra, seg umudam, ɣef tqecwalt-ik, mbla profile deffir-sen.
You · 20 prompts
The web sees · 5 queries
Wagi yerẓa aseɣwen i ɣef tebna internet yettwaḥellsen s tehri, yerna ula d yiwen ur yeḥwaǧ ad t-yesnes s uɛeqda. Amsekyad, s tɣara-s, d win yekksen timagit.
Ur ilaq ara ad yili akka s lekmal. Tuɣzin n timagit d afran i yexdem umsaɣ, yerna aṭas n imsaɣen ad qqimen ṭṭfen deg profile, axaṭer profile d nettat i d ttjara. Maca i tikkelt tamezwarut, afran ur nettwali ara d ashel, yerna win ara yessufɣen amudam adigan, d nettat i ara yefrnen amagnu.
Ayen akka i d-yerran srid ɣer uḥraz amezwaru.
Aseggas aya tili ad iniɣ "assa neɣ wass-nniḍen". Yiwen umudam war isem deg ufella n leaderboard yewwi-yi-d azemz-nni ɣer tallit anda sɛiɣ leɛtimad ugar.
Iɛdaden: Epoch AI (tanezzayt ldin/yeɣlaqen, tarrayin n tazmert n usnerni, aɣmis n tezmar), Artificial Analysis Intelligence Index, tabrat n tuffɣa n Z.ai, isefka n useqdec n OpenRouter, ɣuct 2026.
Iɣbula13
- Ox Alpha on OpenRouter: Free 1M Stealth Model (Aug 2026)
- Zhipu AI shares jump as viral Ox Alpha model revealed as GLM-5.3-Flash on Chinese chips
- GLM-5.2 is probably the most powerful text-only open weights LLM
- GLM-5.3: Scaling with post-training, intuitively explained
- Open vs. closed AI: How behind are open models?
- Open-weight models lag state-of-the-art by around 3 months on average
- Open models lag state-of-the-art closed models by 4 months
- Frontier AI capabilities can be run at home within a year or less
- Pre-training compute efficiency trend
- The performance per dollar of AI chips purchased each quarter has grown by an average of 49% per year
- AI capabilities progress has sped up
- microsoft/phi-4 model card
- Qwen/Qwen3.6-27B model card