I-Reflection Imemetela i-Beam: I-model Lenesisindvo Lesivulekile Lenema-parameter langu-501 billion Lesebentisa Kuphela langu-23 billion Ngesikhatsi Sinye

Imphendvulo lemfisha: Mhla ziyi-5 iMphala 2026, inkampani yaseMelika i-Reflection AI yamemetela i-Beam, i-model yayo yekucala lenesisindvo lesivulekile. Inema-parameter langu-501 billion sewonkhe, kodvwa yakhiwe njenge-model ye-mixture of experts (kuhlanganiswa kwetingcweti), ngako kuphela ema-parameter langu-23 billion lasebentako ku-token ngayinye, futsi iphatsa i-context lesebentako lefika ema-token lasigidzi sinye. Tisindvo titakhishwa ngaphasi kwelilayisensi i-Apache 2.0, levumela kusetjentiswa ekuhwebeni, "ngasekupheleni kwalenyanga", kanye nembiko wetebuchwepheshe, i-model card kanye nemathulusi etinjiniyela. Kuze kube lamuhla awukwati kuyilandza; lokukuphela kwekukhetsa kubhalisela luhlelo lwekufinyelela kusenesikhatsi.

Loku kubalulekile kunoma ngabe nguliphi libhizinisi lelicabanga kusebentisa i-AI kuma-server alo esikhundleni sekutfumela idatha yalo kusevisi ye-cloud lengaphandle kweMbuso. Kodvwa fundza tinombolo letishicilelwe ngekuthula ngaphambi kwekwenta tinhlelo ngato.

Loko i-Reflection Lekumemetele Ngempela

  • Bukhulu: ema-parameter langu-501 billion sewonkhe, langu-23 billion lasebentako ku-token ngayinye, kusetjentiswa kwakheka lokusabalele kwe-mixture of experts.
  • Kucecesha: ema-token langu-23.8 trillion edatha ye-web nemaqoqo edatha lanemalayisensi, kuma-GPU langu-6,144 e-NVIDIA GB300 ngesikhatsi lesingaphansi kwemaviki lamane.
  • Kufundza ngekucinisa (reinforcement learning): emaviki lamane lengetiwe kuma-GPU e-GB300 lacishe abe ngu-10,500, nema-rollout ekucecesha langetulu kwa-100 million.
  • Kusetjentiswa lokuhlosiwe: kubhala ikhodi, kucabanga ngekuhlatiya kanye nemsebenti we-agent, lapho i-model yenta luchungechunge lwetinyatselo ngekwayo.
  • Kulawula kucabanga: sisetjentiswa lesincumako kutsi i-model "icabanga" kangakanani ngaphambi kwekuphendvula, ngako ungakhetsa emkhatsini wemphendvulo lemfisha lesheshako nekuhlatiya lokudzimba.
  • Kusebenta kahle: inkampani itsi i-Beam itfola emaphuzu lacatsaniseka ne-GLM-5.2 kuma-benchmark ekucabanga lokusezingeni leliphezulu, ngaphansi kancane kwayo etafuleni layo, ngesikhatsi isebentisa, ngekulinganisa, emandla ekubala e-inference lancane katsatfu kuya kane. Leyo nombolo ibalwa kusuka kuma-parameter lasebentako nema-token lakhicitwako, hhayi kusuka etindlekweni tekusebentisa letilinganisiwe.

Tinombolo Njengobe Tishicilelwe, Nalabayihlulako

I-Reflection ishicilele lithebula lekucatsanisa nema-model lasikhombisa, lafaka ekhatsi i-GLM 5.2, i-Qwen 3.8 Max ne-Kimi K3. Lena yimiphumela lemikhulu mayelana nalawo lamatsatfu, njengobe kushiwo esimemetelweni:

  • Terminal Bench v2.1 (imisebenti ye-command line): i-Beam 80.1, ngekucatsaniswa na-81.0 ye-GLM 5.2, 86.6 ye-Qwen 3.8 Max kanye na-88.3 ye-Kimi K3.
  • MCP Atlas (kusebentisa emathulusi ngemtsetfo we-MCP): i-Beam 78.7, ngekucatsaniswa na-77.8, 84.5 kanye na-82.3.
  • GPQA Diamond (imibuto yesayensi yetingcweti): i-Beam 90.5, ngekucatsaniswa na-91.2, 92.6 kanye na-93.5.
  • DeepSWE v1.1 (kulungisa tinkinga tangempela te-software): i-Beam 44.4, ngekucatsaniswa na-44.0, 51.0 kanye na-68.0.
  • AIME 2026 (tibalo): i-Beam 97.8 ngekucatsaniswa na-99.2 ye-GLM 5.2.

Kufundza lokwetsembekile kwalelithebula: i-Beam akusiyo i-model levulekile lenemandla kakhulu. I-Kimi K3 iyayihlula kuma-benchmark langu-13 kulangu-14 lapho tombili tinemaphuzu khona (lokukuphela kwalokuhlukile yi-tau3 banking, 38.0 ye-Beam ngekucatsaniswa na-37.1), futsi ngelibanga lelikhulu ekulungiseni ikhodi (68.0 ngekucatsaniswa na-44.4 ku-DeepSWE). Loko i-Reflection lekutsengisako kusebenta kahle: kusebenta lokusedvute ne-GLM-5.2 ngetindleko tekusebentisa letiphansi. Emaphuzu e-Beam abikwa yi-Reflection ngekwayo, kantsi inkampani itsi emaphuzu etimbangi yawatsatsa ku-Artificial Analysis nase-DataCurve. Ematsimba latimele angeke akhone kuphindze ahlole i-Beam kuze kushicilelwe tisindvo nembiko wetebuchwepheshe.

Langu-23 billion Lasebentako Akusho i-server Lencane

Lena yincenye leyengetwa kakhulu kabi mayelana nema-model e-mixture of experts. Linani lema-parameter lasebentako libeka emandla ekubala ku-token ngayinye, lokusho sivinini netindleko temphendvulo ngayinye. Kodvwa imemori isafanele iphatse yonkhe i-model. Ngetibalo letilula, ema-parameter langu-501 billion adzinga cishe 500 GB yemamemori ye-GPU nga-8 bits ngalinye, noma cishe 1 TB nga-16 bits, kwetisindvo todvwa, ngaphambi kwe-cache ye-context nebasebentisi labasebentisa ngesikhatsi sinye. I-Reflection ayikasho kutsi itakhipha tisindvo ngaliphi lizinga lekunemba. Ukhuluma nge-server lenema-GPU lamanyenti lasezingeni leliphezulu, hhayi ngemshini we-desktop. I-Reflection ayikashicileli tidzingo letisemtsetfweni tekusebentisa noma tinguculo letincishisiwe, ngako balindze ngaphambi kwekulinganisa sabelomali.

Loku Kusho Kutsini Emabhizinisini aseSaudi

Lusito lwangempela lwanoma ngabe nguyiphi i-model lenesisindvo lesivulekile lusendzaweni lapho isebenta khona. Nangabe i-model isebenta kuma-server akho noma esikhungweni sedatha ngekhatsi kweMbuso, idatha yemakhasimende, tivumelwano nema-invoice kuhlala ngaphasi kwekulawula kwakho. Loko kwenta kube melula kulandzela uMtsetfo Wekuvikelwa Kwedatha Yemuntfu (PDPL), lowengamelwa Siphatsimandla SaseSaudi Sedatha Nebunhlakanipho Lobentiwe (SDAIA), ikakhulukati imitsetfo yawo mayelana nekudlulisela idatha ngaphandle kweMbuso.

  • Ema-agent ekubhala ikhodi langekhatsi: licembu letebuchwepheshe lelifuna umsiti wekubhala ikhodi ngaphandle kwekutfumela ikhodi yalo yemtfombo kusevisi yangaphandle.
  • Kwenta tinchubo titenteke ngekwato: i-agent lefundza ticelo futsi ibite emathulusi etinhlelo tangekhatsi, lapho emaphuzu e-MCP Atlas abaluleke kakhulu kunema-benchmark etibalo.
  • Kuhlatiya emadokhumenti ladzimba: i-context lenema-token lasigidzi sinye ikuvumela kutsi ufundze tivumelwano letiphelele noma ema-log ngesikhatsi sinye, nangabe uhlola likhwalithi yekucondza kumafayela akho.

Sicwayiso lesibalulekile: simemetelo asibiti tilwimi letisekelwako, futsi asikhulumi nga-Arabhu. Umphumela lowodvwa wetilwimi letinyenti yi-SWE-Bench Multilingual nga-78.0, futsi loko kusho tilwimi tekuhlela ema-computer, hhayi tilwimi tebantfu. Ungacabangi kutsi likhwalithi yesi-Arabhu ilungile ngaphambi kwekutsi uyihlole.

Indlela Yekusondzela Kuyo, Ngesinyatselo Ngesinyatselo

  • Ungashintji luhlelo lwakho lamuhla: i-model ayitfolakali kutsi ilandzwe, futsi tinombolo atikabuyeketwa ngendlela lecinile.
  • Lungisa lisetjana lekuhlola lelivela emsebentini wakho: imisebenti yangempela lesuka ku-50 kuya ku-100, njengemilayeto yemakhasimende ngesi-Arabhu, ticelo netincwadzi, nemphendvulo lelungile kuyo yonkhe.
  • Catsanisa nangabe tisindvo tifika: i-Beam ngekucatsaniswa ne-model levulekile losuvele uyicabanga, njenge-Kimi K3, ngekunemba, kubambeka kanye netindleko te-server.
  • Bala tindleko letiphelele: i-server, gesi, kusebenta nekunakekela, ngekucatsaniswa netindleko te-API ye-cloud ngelinani lelifanako leticelo.
  • Yakha sendlalelo sekushintja: yenta luhlelo lwakho lubite i-model ngekusebentisa i-interface yinye, kute ukhone kushintja emkhatsini wema-model ngaphandle kwekwakha kabusha i-application.

Le yindlela lesisebenta ngayo e-Origami nasakha tinhlelo te-AI temakhasimende etfu: sihlola i-model ngedatha yelibhizinisi ngekwalo, futsi sakha luhlelo ngendlela yekutsi i-model ibe yincenye lengashintjwa, kute simemetelo ngasinye lesisha singaphendvuki sibe yiphrojekthi yekwakha kabusha. Nangabe idatha yakho ibucayi, fundza umhlahlandlela wetfu mayelana noMtsetfo Wekuvikelwa Kwedatha Yemuntfu ngaphambi kwekukhetsa kutsi uyisebentisela kuphi.

Imitfombo