Amamodeli we-Open 7 Open Open Ai Amakhodi Ai Uphuthelwa

Isithombe nguMlobi

Okusobala Ukuqalisa

Iningi labantu abasebenzisa abasizi be-Artificial Intelligence (AI) Abasizi be-Coding namuhla bathembela kumathuluzi asuselwa efwini afana Ikhodi yeClaude, Github Copilot, Isikhohlisinabanye. Banamandla, akungabazeki. Kepha kukhona ukucasha okukhulu kokuhweba okukhulu emehlweni asobala: Ikhodi yakho kufanele ithunyelwe kumaseva womunye umuntu ukuze la mathuluzi asebenze.

Lokho kusho ukuthi wonke umsebenzi, lonke ukhiye wokuhlela isikhombi sohlelo (API), zonke izinqumo zokwakha zangaphakathi ziyadluliselwa -Obubizin, Vulainoma omunye umhlinzeki ngaphambi kokuthola impendulo yakho emuva. Futhi noma bethembisa ubumfihlo, amaqembu amaningi avele athathe lelo ngozi. Ikakhulukazi uma usebenza nge:

• Amakhodi okuphathelene noma okuyimfihlo
• Izinhlelo zamakhasimende ezibhizinisi
• Ukucwaninga noma ukulayishwa kukahulumeni
• Noma yini ngaphansi kwesivumelwano esingadaluli (NDA)

Yilapho Amamodeli we-Local, Ovulekile Womthombo Ovulekile Shintsha umdlalo.

Ukugijima imodeli yakho ye-AI kukunika indawo yokulawula, ubumfihlo, nokuvikeleka. Ayikho i-Code eshiya umshini wakho. Azikho izingodo zangaphandle. Cha “wethembe.” Futhi ngaphezulu kwalokho, uma usunehardware enekhono, ungagcina izinkulungwane ku-API kanye nezindleko zokubhalisa.

Kulesi sihloko, sizohamba ngamamodeli ama-AI avulekile avulekile avulekile aqhelelane ngamaphuzu aqheleleni ama-benchmark amakhodi futhi abe ngamanye izindlela zangempela kumathuluzi okuphathelene.

Uma ufuna inguqulo emfushane, skrolela ezansi kwetafula lokuqhathanisa elisheshayo lawo wonke amamodeli ayisikhombisa.

Okusobala 1. Kimi-K2-Ukucabanga NgeMoonsonT AI

Kimi-K2-Ukucabangakwakhiwa ngu Moonshot aiiyimodeli yokucabanga ethuthukisiwe evulekile eyenziwe njenge-ejenti yokusebenzisa ithuluzi elinezizathu zesinyathelo ngesinyathelo ngesinyathelo ngenkathi kuvuswa imisebenzi nezinsizakalo. Igcina i-ejensi ende ezinzile ngaphesheya kwezingu-200 kuya ku-300 ithuluzi lokulandelana elilandelana – ukuthuthuka okukhulu ngaphezulu kwe-30 kuya ku-50-stap drift ebonwe ezinhlelweni zangaphambilini. Lokhu kunika amandla ukugeleza komsebenzi ozimele ekucwaningeni, ukufaka amakhodi, nokubhala.

I-Architactaly, i-K2 Ecabangayo ibonisa imodeli enamapharamitha ayizigidigidi eziyi-1, lapho ama-32 billion asebenza khona. Kubandakanya ochwepheshe abangama-384 (abakhethwe abangu-8 ngakunye futhi babelwane nge-1), izingqimba ezingama-61 (ezinezendlalelo ezi-1 ezinobukhulu), nobukhulu bezinhloko ezingama-7,168 ngamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-7,168 anamakhanda angama-74. Isebenzisa ukunakwa kwe-MLA kanye nokwenza kusebenze. Imodeli isekela iwindi lomongo lamathokheni angama-256,000 futhi inesilulumagama sabangu-160,000. Kuyimodeli yendabuko ye-Int4 esebenzisa ukuqeqeshwa kwe-post-okwaziwa (qat), okuholela cishe ku-2 × isivinini kumodi ephansi ye-latency ngenkathi kunciphisa ukusetshenziswa kwememori ye-GPU.

Isithombe nguMlobi

Ezivivinyweni ze-benchmark, ukucabanga kwe-K2 kufinyelela imiphumela ehlaba umxhwele, ikakhulukazi ezindaweni lapho ukucabanga okuhlala isikhathi eside nokusebenzisa ithuluzi kubucayi. Ukusebenza kwamakhodi kulinganiselwe kahle, ngezikolo ezifana ne-SWE-Bench ziqinisekiswe ngo-71.3, ama-multi-sye at 41.9, scicode ngo-44.8, ne-terminal-bench ngo-47.1. Ukusebenza kwalo kokuma kubonakala ku-LiveCodebelch V6, lapho kwathola khona amaphuzu angama-83, lapho kukhombisa amandla athile ekuhambeni kwezilimi eziningi kanye nokusebenza kwama-agentic.

Okusobala 2. I-MiniMax-M2 nge M2

Le khasi Minax-m2 Ukuhlelwa kabusha kokusebenza kahle kokuhamba komsebenzi osuselwa ku-Agent-based. Kuyingxubevange ye-compact, esheshayo, ebizayo yemodeli (MOE) ye-SOLECTS equkethe ingqikithi yamapharamitha ayizigidi ezingama-230, enamabhiliyoni ayi-10 kuphela acushiwe ngethokheni ngalinye. Ngokuhambisa ochwepheshe abafanelekile, i-Minimax-M2 ifinyelela ukusebenza kokusetshenziswa kwamathuluzi okugcina ngokujwayelekile okuhambisana namamodeli amakhulu ngenkathi kunciphisa i-latency, izindleko, kanye nokusebenzisa inkumbulo. Lokhu kwenza kube fanele ama-ejenti asebenzayo kanye nesampula ye-batted.

Yakhelwe amakhodi we-Elite Amakhodi Nemisebenzi ye-Agent ngaphandle kokuyekethisa ubuhlakani obujwayelekile, igxile ohlelweni → isenzo → xwafesa. Lezi ziloops zihlala zisabela ngenxa yezinyawo ezisebenza ngezigidi eziyizinkulungwane eziyishumi.

Isithombe nguMlobi

Kuma-benchmark amakhodi we-real-world kanye nama-agent, imiphumela ebikiwe ikhombisa ukusebenza okuqinile okusebenzayo: I-SWE-BENCH ifaka amaphuzu angama-69.4, amabhentshi amaningi angu-36.2, i-terminal-ebhentshini 46.3, kanye ne-artifartsbench 66.8. Kwama-web kanye nama-ejenti wokucwaninga, izikolo zimi kanjena: isiphequluli 44 (ngesikolo esingu-48,5 ngesiShayina), Gaia (Umbhalo) 75.2, i-HECH-SETSOMS 72,8, kanye neFinsearchComp-Global 65.5.

Okusobala 3. I-GPT-OSS-120B ka-Opena

I-GPT-OSS-120B iyimodeli ye-moe esivulekile eyenzelwe ukusetshenziswa kokukhiqizwa ngokukhiqizwa ngokujwayelekile, imithwalo ephezulu yokubonisana. Ilungiselelwe ukusebenza ku-80GB GPU futhi ifaka amapharamitha ayizigidi eziyizinkulungwane ezingama-117, ngamapharamitha ayizigidi eziyizinkulungwane ezingama-5.1 ngethokheni ngalinye.

Amakhono asemqoka we-GPT-OSS-120B afaka amazinga okubonisana avumelanayo (aphansi, aphakathi nendawo), ukufinyelela okugcwele kwe-chain-of-eventuct), ukuphequlula, ukuhlanganiswa kwe-python, nokusekelwa okugcwele okuhle, kanye nokusekelwa okugcwele okuhle. Ngokwengeziwe, imodeli encane yomngane omncane, i-GPT-OSS-120B, iyatholakala kubasebenzisi abadinga i-latency ephansi kanye nezinhlelo zokusebenza zendawo ezilungiselelwe.

Isithombe nguMlobi

Ekubhekeni kwangaphandle, amazinga we-GPT-OSS-120B njengemodeli yesithathu ephezulu kakhulu ku- Inkomba yokuhlaziya ubuhlakani. Ibonisa okunye ukusebenza okuhle kakhulu nesivinini esihlobene nosayizi wayo, ngokususelwa ekuqhathaniseni okuzenzakalelayo kwekhwalithi yekhwalithi, ijubane lokukhipha, kanye ne-latency.

I-GPT-OSS-120B ihluza i-O3-Mini nemidlalo noma idlula amakhono we-O4-Mini ezindaweni ezinjengemininingwane yokuncintisana (amakhodi wokuxazulula izixazululo (i-TAUBU, HLE), kanye nokusetshenziswa kwamathuluzi (i-Taubency). Ngaphezu kwalokho, idlula i-O4-Mini ekuhlolweni kwezempilo (impilo yezempilo) kanye nezibalo zokuncintisana (AIME 2024 no-2025).

Okusobala 4. I-Deepseek-V3.2-Exp nge-Deepseek AI

Deepseek-v3.2-exp yisinyathelo esiphakathi sokuhlola esizukulwaneni esilandelayo se Deepseek aiukwakhiwa kwezakhiwo. Yakha phezu kwe-v3.1-terminus futhi yazisa ukunakwa okujulile (i-DSA), indlela yokunakwa ehlelwe kahle eyenzelwe ukuthuthukisa ukusebenza kahle kokuqeqeshwa kanye nokuhlolwa kwezimo ezingokoqobo.

Ukugxila okuyisisekelo kwalokhu kukhululwa ukuqinisekisa ukuzuza okusebenzayo kokulandelana okunwetshiwe ngenkathi kugcinwa isenzo esiqinile semodeli. Ukuhlukanisa umthelela we-DSA, ukucushwa kokuqeqeshwa kuhambisana ngamabomu nalezo ze-V3.1. Imiphumela ikhombisa ukuthi ikhwalithi yemiphumela ihlala icishe ifane.

Isithombe nguMlobi

Kuwo wonke amabhentshi asekhaya, i-v3.2-exp yenza okufanayo ku-v3.1-terminus, ngamashifu amancane ekusebenzeni: I-MMLU-pro esondele ku-85.0, ibonakale eduze kwe-80.7 uma iqhathaniswa no-80.7). Ngokwengeziwe, kukhona izinzuzo ku-aime 2025 (89.3 kuqhathaniswa nama-88.4) kanye namakhodi (2121 kuqhathaniswa no-2046).

Okusobala I-5. Glm-4.6 nge-z.ai

Qhathanisa ne-glm-4.5, Glm-4.6 Inwebisa iwindi lomongo kusuka ku-128k kuya ku-200k amathokheni. Lokhu kuthuthukiswa kuvumela ukugeleza komsebenzi okuyinkimbinkimbi futhi okude kakhulu ngaphandle kokulahlekelwa yithrekhi yolwazi.

I-GLM-4.6 futhi inikeza ukusebenza kwamakhodi aphezulu, ukuzuza izikolo eziphakeme kumabhentshi wekhodi kanye nokuletha imiphumela yezwe yangempela yamathuluzi afana nekhodi yeClaude, Uzimise, Ikhodi ye-Roone-Kilo Code, kufaka phakathi isizukulwane esiphikisayo esichithiwe.

Isithombe nguMlobi

Ngokwengeziwe, i-GLM-4.6 yethula amandla okubonisana athuthukile ngokusetshenziswa kwamathuluzi ngesikhathi sokutholwa, okuqinisa ukusebenza kwayo okuphelele. Le nguqulo ifaka ama-ejenti anekhono ngokwengeziwe ngokusetshenziswa kwamathuluzi okuthuthukile kanye nokusebenza kwe-ejenti, kanye nokuhlanganiswa okunzima ngaphakathi kwezinhlaka ze-ejenti.

Kuwo wonke amabhentshi ayisishiyagalombili asemdlalweni akhombisa ama-ejenti, abonisana, kanye nokufaka amakhodi, amakhodi-4.6 akhombisa ukuthuthuka okucacile ngaphezulu kwe-GLM-4.5 futhi agcina izinzuzo zokuncintisana ngokuqhathaniswa namamodeli anjenge-deedeek-v3.1-terminus noClaude Sonnet 4.

Okusobala I-6. Qwen3-355B-A22B-Emfundweni

Qwen3-235B-A22B-Stime-2507 ukungacabangi okungacabangi kwe I-Alibaba CloudImodeli ye-Flagship ye-Flagship, eyenzelwe ukusetshenziswa okungokoqobo ngaphandle kokuveza inqubo yayo yokubonisana. Inikeza ukuthuthukiswa okukhulu kumandla ajwayelekile, kufaka phakathi imiyalo elandelayo, ukucabanga okunengqondo, izibalo, isayensi, ukufaka amakhodi, nokusetshenziswa kwamathuluzi. Ngokwengeziwe, kwenze intuthuko enkulu ngolwazi olude lomsila wazo zonke izilimi eziningi futhi kukhombisa ukuhambisana okuthuthukile nokuthandwa ngumsebenzisi kwemisebenzi evuthiwe nemisebenzi evulelekile.

Njengemodeli yokungacabangi, inhloso yayo eyinhloko ukukhiqiza izimpendulo eziqondile kunokuba unikeze izindlela zokubonisana, ugxile ekusizeni nasekuthendeni okuphezulu kombhalo wekhwalithi yansuku zonke.

Isithombe nguMlobi

Ekuhlolweni komphakathi okuhlobene nama-ejenti, ukubonisana, nokufaka amakhodi, kukhombisile ukuthuthuka okucacile kokukhishwa kwangaphambilini futhi kugcina umngcele wokuncintisana ngaphezulu kwamamodeli avulekile namamodeli okuphathelene (isib.

Okusobala 7

I-Apriel-1.5-15b-Imcabangi IS Servicenow aiImodeli yokubonisana ye-multimodal evela ku-Apriel South Model Model (SLM) Series. Kwethula amakhono okubonisana kwezithombe ngaphezu kwemodeli yombhalo wangaphambilini, egqamisa irejimeni eliphakathi nendawo lokuqeqeshwa okubandakanya ukuqhubeka okuqhubekayo kombhalo kanye nezithombe, kulandelwe yi-SFT-kuphela yokuqondisa okuhle (i-RL). Naphezu kosayizi wayo ocwengekileyo wamapharamitha ayizigidi eziyizinkulungwane eziyi-15, okuvumela ukuthi isebenze ku-GPU eyodwa, inezinkinga zobude obubikwe cishe amathokheni ayi-131,000. Le modeli ihlose ukusebenza kanye nokusebenza kahle okuqhathaniswa namamodeli amakhulu amakhulu, cishe amahlandla ayishumi, ikakhulukazi emisebenzini yokubonisana.

Isithombe nguMlobi

Kumabhentshini omphakathi, i-apriel-1.5-15b-mbongiler ifinyelela amaphuzu angama-52 kwizinkomba zobuhlakani bokuhlaziya ubuhlakani, okwenza kuncintisene namamodeli afana ne-Deepseek-R1-0528 ne-Gemini-Flash. Kuthiwa okungenani yingxenye eyodwa yeshumi kunoma yikuphi ukubheja imodeli engenhla 50. Ngaphezu kwalokho, kukhombisa ukusebenza okuqinile njenge-ejenti yebhizinisi, amagoli angama-68 ku-Ifbench.

Okusobala Ukufingqa itafula

Nasi isifinyezo semodeli yomthombo ovulekile ngecala lakho elithile lokusebenzisa:

Abid Awan Awan (@ 1Abidaliawan) ungumqeqeshi oqinisekisiwe wesosayensi othanda amamodeli wokufunda umshini. Njengamanje, ugxile ekudalweni kokuqukethwe nangokubhala amabhulogi ezobuchwepheshe ekufundeni komshini kanye nobuchwepheshe besayensi yedatha. I-Avid ibamba iziqu ze-master ekuphathweni kwezobuchwepheshe kanye neziqu ze-bachelor ku-telecommunication Engineering. Umbono wakhe ukwakha umkhiqizo we-AI usebenzisa inethiwekhi ye-graph neural for abafundi abalwela ukugula kwengqondo.