Hlangana ne-OpenJarvis: Uhlaka Lwendawo Lokuqala Lwama-ejenti e-AI yomuntu siqu akudivayisi onamathuluzi, inkumbulo, nokufunda.

Abacwaningi baseStanford University kanye naseLambda Labs, bashicilele iphepha lokucwaninga le-OpenJarvis, uhlaka lomthombo ovulekile olusebenzisa i-inference, ama-ejenti, inkumbulo, nokufunda ngokuphelele kudivayisi.

Amamodeli anesisindo esivulekile acushwa ngomhlaba we-OpenJarvis phakathi kwamaphesenti angu-3.2 wemodeli yamafu ehamba phambili ngokwesilinganiso, cishe ngezindleko eziphansi ze-API engu-800× embuzweni ngamunye futhi cishe u-4× ukubambezeleka okuphansi ngaphansi kwephrothokholi yebhentshimakhi yocwaningo. Lo msebenzi wocwaningo wakhela phezu kwethimba labacwaningi ngaphambilini Intelligence Ngamunye Watt Ucwaningo, olubike ukuthi amamodeli endawo asevele ephethe u-88.7% wengxoxo eyodwa kanye nemibuzo yokucabanga ngesikhathi sokulinda okusebenzisanayo, nokusebenza kahle kobuhlakani kuthuthukisa i-5.3 × kusuka ku-2023 kuya ku-2025.

Ukubuka konke kwemodeli nokufinyelela

I-OpenJarvis akuyona imodeli eyodwa. Kuwuhlaka oluhlanganisa noma iyiphi imodeli esekelwayo enestakhi somenzeli olungisekayo, esihlolwa kumamodeli endawo angu-11 avela emindenini emine.

I-Architecture: Ama-Primitives amahlanu kanye ne-Spec

I-OpenJarvis ibolisa isistimu yomuntu siqu ye-AI ibe yiziqalo ezinhlanu ezithayiphiwe, ezakhiwe ngento eyodwa eyisimemezelo ebizwa ngokuthi spec.

• Ubuhlakani – imodeli, izisindo, imingcele yesizukulwane, nefomethi yokulinganisa.
• Injini — isikhathi sokusebenza sokucatshangelwa (i-Ollama, i-vLLM, i-SGlang, njll.), i-batching, izilungiselelo ze-KV-cache, kanye nendlela yehadiwe.
• Ama-ejenti – iluphu yokucabanga (i-ReAct noma i-CodeAct), ukwaziswa kwesistimu, inqubomgomo yokusetshenziswa kwamathuluzi, nemikhawulo yokujika.
• Amathuluzi Nenkumbulo – ukuxhumana kwangaphandle, okungemuva kokubuyisa, izixhumi zedatha ezingama-25+, neziteshi zemiyalezo ezingama-32+, ezinokwesekwa komdabu kwe-MCP kanye nezingemuva ezishintshashintshayo zememori.
• Ukufunda – i-optimizer ebuyekeza i-spec kusuka kumathrekhi. Lesi sikhala samukela i-LoRA, i-DSPy, i-GEPA, noma i-LLM-guided spec search search.

I-primitive ngayinye ingashintshwa ngokuzimela, futhi i-spec ihlela zonke ezinhlanu zibe ifayela le-TOML. Ukucaciswa okubili kungabelana nge-ejenti efanayo nokucushwa kwethuluzi futhi kuhluke kuphela ngemodeli nenjini, ngakho ukuziphatha okufanayo kusebenza ku-Mac Mini kanye nendawo yokusebenza ngaphandle kokwaziswa kokubhala kabusha.

Usesho oluqondisiwe lwe-LLM umnikelo wesibili. Ukusebenzisana kwasendaweni namafu: imodeli yefu eliwumngcele isebenza njengothisha ngesikhathi sokusesha, ukulandela umkhondo, ukuhlonza amaqoqo okuhluleka, nokuhlongoza ukuhlela kuyo yonke i-Intelligence, Injini, Ama-ejenti, kanye Namathuluzi & Nenkumbulo. Ukuhlela kwamukelwa kuphela uma kuthuthukisa iqoqo lokuhluleka okuqondiwe ngaphandle kokubangela ukuhlehla okunengqondo kwenye indawo – ithimba locwaningo libiza lokhu ngokuthi Isango (ukubekezelelana okuzenzakalelayo 1%). Ukucaciswa okuthuthukisiwe bese kusebenza ngokuphelele kudivayisi ngesikhathi se-inference, ngezingcingo zamafu eziyiziro. Uthisha usetshenziswa kuphela ngesikhathi sokucinga; ngemibuzo eyi-100 ngosuku, izindleko zikathisha ezincishisiwe ziwela ngaphansi kuka-$0.001 ngombuzo ngamunye phakathi nezinyanga eziyisithupha.

Umsebenzi wangaphambili (i-GEPA, i-DSPy, i-LoRA) ithuthukisa okwakudala okukodwa ngesikhathi, futhi izilungiseleli ezisheshayo zizodwa zithola cishe u-5 pp wegebe lendawo yamafu. Ukusesha okuqondisiwe okuqondiswa yi-LLM kuthola kabusha okungu-13–32 pp ngoba kuhlela kuwo wonke ama-primitives ngokuhlanganyela, ngezindleko eziphansi zokuthuthukisa ezingu-7–11× kunezisekelo eziyisisekelo esisodwa. Isikhala sokunyakaza sakudala ezine sinikela ngo-5.5–16.5 pp, futhi umhlongozi we-LLM wengeza cishe u-10 pp ngokwesilinganiso phezu kokusesha kokuziphendukela kwemvelo endaweni efanayo yokunyakaza.

Amakhono & Ukusebenza

I-OpenJarvis yahlolwa kuwo wonke amabhentshimakhi angu-8 ahlanganisa imisebenzi engu-508: ukushaya ithuluzi (ToolCall-15), ukugeleza komsebenzi we-agent (PinchBench), ukufaka amakhodi (LiveCodeBench), isevisi yamakhasimende (τ-Bench V2, τ²-Bench Telecom), usizo olujwayelekile (GAIA), kanye nocwaningo olujulile (LiveResearchBench, DeepRese).

Ukuhlolwa kokushintshana: Ukushintsha imodeli yefu ehlosiwe nge-Qwen3.5-9B kuzinhlaka ezikhona kakade (OpenClaw, Hermes Agent) kwehlisa ukunemba ngo-25–39 pp. Ngemodeli efanayo ngaphansi kwe-OpenJarvis spec, ukwehla kwensalela kuncipha kuya ku-5.6–16.5 pp — kubuyisela ukulahleka kokuphatheka okungu-56–77.

Umngcele wokunemba: Imodeli yendawo eyodwa engcono kakhulu, i-Qwen3.5-122B, ifinyelela ku-80.3% ukunemba okumaphakathi uma iqhathaniswa ne-Claude Opus 4.6 ku-83.5% – igebe le-3.2 pp. Ukucaciswa kwendawo kufana noma kudlule ifu kumabhentshimakhi angu-4 kwangu-8: I-ToolCall-15, i-PinchBench, i-LiveCodeBench, ne-τ-Bench V2.

Izindleko nokubambezeleka: Ukucushwa kwendawo kwakha umngcele wokunemba nokusebenza kahle. I-Qwen3.5-122B iletha u-80.3% wayo cishe engxenyeni eyinkulungwane yesenti ngombuzo ngamunye, uma iqhathaniswa no-$0.009 ngombuzo ngamunye we-Claude Opus 4.6 – inzuzo ecishe ibe ngu-800× emaphethelweni we-API-izindleko. Ukubambezeleka kokuphela kwehla cishe ngo-4 × kumthwalo wemisebenzi ye-ejenti, nakuba amanothi ephepha aphawula ngokudubula okukodwa angavuna ukusebenzela amafu.

Sesha izinzuzo: Ukusesha okuqondisiwe okuqondiswa yi-LLM kuthuthukisa isitshudeni se-Qwen3.5-9B siye ku-100% ku-PinchBench, 83% ku-LiveCodeBench, kanye no-91% ku-LiveResearchBench. Kuhlelo oluphelele lwamabhentshimakhi ayisishiyagalombili, izinzuzo ezimaphakathi ngemodeli yomfundi ngamunye zisukela ku-13.1 kuye ku-31.5 pp. Ababhali babika ukuthi lezi zinzuzo zisinda ekuhlolweni kwazo kokuqina (okuhlukile kwesisindo somvuzo, ukuhluka kwembewu yosesho, nokuqaliswa kabusha okungahleliwe).

Isetshenziswa kanjani

Ukufaka umyalo owodwa. Ku-macOS, Linux, noma i-WSL2:

curl -fsSL  | bash

Abasebenzisi beWindows basebenzisa iskripthi esifanayo se-PowerShell (irm … | iex). Izinhlinzeko zesifaki uvindawo ebonakalayo ye-Python, i-Ollama, kanye nemodeli yokuqalisa cishe imizuzu emithathu ku-broadband. I-GUI yedeskithophu ithunyelwa njenge- .dmg, .exe, .deb, .rpmnoma .AppImage kusuka ekhasini lokukhishwa.

Ngemva kokufaka, jarvis iqala iseshini yengxoxo. Ukusetha kuqala kuhlanganisa ukuhamba komsebenzi okuvamile:

jarvis init --preset morning-digest-mac    # daily briefing with TTS
jarvis init --preset deep-research         # multi-hop research with citations
jarvis init --preset code-assistant        # agent with code execution and shell access
jarvis init --preset scheduled-monitor     # stateful agent on a schedule

Uhlaka luhamba nama-ejenti angu-8 akhelwe ngaphakathi kuzo zonke izindlela ezintathu zokwenza – okudingekayo, okuhleliwe, nokuqhubekayo. Ixhumeka emithonjeni yedatha engu-25+ (i-Gmail, Ikhalenda, i-iMessage, i-Notion, i-Obsidian, i-Slack, i-GitHub, neminye) futhi idalula abenzeli abangaphezu kweziteshi zemiyalezo ezingu-32+ (i-WhatsApp, iTelegram, i-Discord, i-iMessage, iSignal, nezinye).

Amakhono angangeniswa evela kumakhathalogi angaphandle – angaba ngu-150 avela ku-Hermes Agent kanye namakhono omphakathi angaba ngu-13,700 avela ku-OpenClaw – konke kulandelwa ukucaciswa kwe-agentkills.io. A jarvis optimize skills --policy dspy umyalo uyawacwenga ukusuka kumlando wokulandelela wendawo.

Isichazi Esibonakalayo sikaMarktechpost

curl -fsSL  | bash

Okuthathwayo Okubalulekile

• I-OpenJarvis isebenzisa i-inference, ama-ejenti, inkumbulo, nokufunda ngokugcwele kudivayisi, ifika ngaphakathi kwe-3.2 pp yemodeli yefu engcono kakhulu ~ 800× izindleko eziphansi ze-API kanye ~ 4× ukubambezeleka okuphansi.
• “I-spec” ethayiphiwe ibolisa isitaki sibe yizinto zokuqala ezinhlanu ezishintshwayo – Ubuhlakani, Injini, Ama-ejenti, Amathuluzi Nenkumbulo, kanye nokufunda – kuhlelwe ku-TOML ephathekayo.
• Usesho lwe-LLM oluqondiswa yi-LLM lusebenzisa imodeli yefu elisemngceleni njengothisha wesikhathi sokusesha ukuze kubuyiselwe u-13–32 pp wegebe le-cloud–lendawo ngezindleko eziphansi zokulungiselela ezingu-7–11×, bese lusebenza endaweni ngaphandle kwezingcingo zamafu.
• Ukucaciswa kwendawo kufana noma kudlule ifu kumabhentshimakhi angu-4 kwangu-8 (ToolCall-15, PinchBench, LiveCodeBench, τ-Bench V2); igebe elisele ligxile ekucabangeni- kanye nemisebenzi yocwaningo-enzima.

Hlola Iphepha futhi I-Repo. Futhi, zizwe ukhululekile ukusilandela Twitter futhi ungakhohlwa ukujoyina wethu 150k+ ML SubReddit futhi Bhalisela ku Iphephandaba lethu. Linda! ukutelegram? manje ungasijoyina kuthelegramu futhi.

Udinga ukusebenzisana nathi ekuthuthukiseni i-GitHub Repo yakho NOMA Ikhasi Lobuso Lokugona NOMA Ukukhishwa Komkhiqizo NOMA I-Webinar njll.? Xhuma nathi