Startlux-Decision Uyasebenza

I-StartLux-Decision iwumndeni wamamodeli okwenza izinqumo owethulwe yi-StartLux Labs, ohlanganisa amamodeli amahlanu aminyene asukela ku-0.8B kuya ku-27B kanye nemodeli ye-mixture-of-experts engu-35B-A3B. I-StartLux-Decision yamukela isimo esiwumbhalo, i-JSON noma isithombe. Ezinkingeni ezifana nokukhetha, yebo noma cha, kanye nokunikeza amaphuzu, ukubikezela okukodwa kwe-forward pass kubuyisa amathuba enketho ngayinye, angasetshenziswa njengokuzethemba. Amamodeli asekela ngokwemvelo umongo omude wama-token angu-256K, anokubambezeleka okulinganiswa ngama-millisecond, futhi ahambisana nefomethi yeklayenti le-TypeSafe ye-Jev.

Ukwenza izinqumo okuhleliwe: kusekela imibuzo yokukhetha, yebo noma cha, kanye nokunikeza amaphuzu, futhi umbuzo ngamunye ubuyisa ukusatshalaliswa kwamathuba ezinketho.
Okokufaka okunemodi eziningi: isimo singaba umbhalo, i-JSON noma isithombe (kubuyekezwe ngo-2026-10-03); ubufakazi nezithombe kungaba yingxenye yesicelo.
Umongo omude kakhulu: wonke amasayizi amamodeli asekela ngokwemvelo ubude be-prompt obungama-token angu-262,144 (256K), futhi izimo ezinde zifundwa kanye ngezingxenye.
Okukhipha ukuzethemba: amathuba asetshenziswa ngokuqondile njengokuzethemba, okulungele izinqubo ezizenzakalelayo ezidinga ukulinganiswa kobungozi.
Ukuphuma okuningi nokubambezeleka okuphansi: imibuzo emifushane ithola impendulo nge-forward pass eyodwa emizuzwini embalwa nje yama-millisecond, futhi kusekelwa i-batched path nokusheshiswa kwe-CUDA Graphs.
Ukusetshenziswa kuzo zonke izinkundla: kuhlinzekwa i-GGUF quantization (BF16/Q8_0/Q4_K_M) ye-llama.cpp, futhi i-MLX backend iyatholakala ku-Apple Silicon.
Izinhlelo zemidlalo nama-agent ahlakaniphile: sekuqinisekisiwe ezimweni zokwenza izinqumo ngesikhathi sangempela ezifana nokuthenga kuwebhu, i-chess, i-Super Mario, i-StarCraft II ne-Doom.
Ukuhambisana ne-ecosystem ye-Jev: izicelo nezimpendulo zisebenzisa ifomethi ye-TypeSafe /v1/systemone, ngakho amaklayenti abhalelwe i-Jev angaxhumeka ngaphandle komthungo.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI omthombo ovulekile
Ukufunda impendulo nge-forward pass eyodwa: imodeli inikeza izinketho izimpawu zezinhlamvu, bese ifunda ngokuqondile amathuba ezindaweni zezinhlamvu zezinketho ngemva kwe-forward pass. Yonke imibuzo esicelweni iqedwa ngokuhambisana ku-forward pass eyodwa.
Ukwakhiwa kwe-linear attention: kusetshenziswa ama-kernel ashesha we-flash-linear-attention kanye ne-causal-conv1d, okunikeza ukubambezeleka okuphansi kakhulu ezicelweni ezimfushane. Iseva yenqaba ukuqala kuma-GPU angenawo la ma-kernel ukuze kuqinisekiswe ukusebenza.
Ukuphindaphinda kwamamatrix okuqoqwe ku-MoE: imodeli ye-mixture-of-experts engu-35B-A3B ivula cishe amapharamitha angu-3B kuphela ku-token ngayinye. Ama-expert asebenza nge-grouped matmul futhi ambozwe yi-CUDA Graphs, ngakho ukubambezeleka kucishe kube uhhafu kwalokho kwemodeli eminyene engu-27B.
Ukuhlukanisa nokuphindaphinda kokokufaka okude: isimo esingu-256K sifundwa kanye ngezingxenye, bese umbuzo ngamunye esicelweni ubalwa kusuka egatsheni lesethulo sawo, kugwenywe ukufaka ikhodi okuphindaphindiwe.
Ukulondolozwa kokunemba kwamathuba nge-quantization: i-BF16 ne-Q8_0 ye-GGUF inikeza izinqumo ezifana ngokuphelele nezisindo zoqobo kumaphrojekthi angu-99–100% e-JevBench, okubonisa ukuthi ikhono lokwenza izinqumo liqinile kakhulu ku-quantization.

Landa izisindo: sebenzisa i-hf download startlux-models/StartLux-Decision-4B --local-dir StartLux-Decision-4B ukuze ulande ku-Hugging Face.
Faka okuncikile: ngena enkombeni yemodeli bese usebenzisa i-pip install -r requirements.txt.
Qinisekisa ama-kernel: sebenzisa i-python -m startlux_decision.check StartLux-Decision-4B; kufanele kukhishwe “fast kernels: active”.
Qalisa iseva: sebenzisa i-python -m startlux_decision.server --model StartLux-Decision-4B --port 8090 ukuze uqale isevisi yasendaweni yokwenza inference.
Thumela isicelo: thumela i-JSON ku-localhost:8090/v1/systemone, equkethe state kanye ne-questions zohlobo lwe-choice / noul / score.
Thola imiphumela: impendulo ibuyisa ukusatshalaliswa kwamathuba ezinketho zombuzo ngamunye, okungasetshenziswa ngokuqondile njengokuzethemba kwesinqumo.
Indlela ye-GGUF (ongayikhetha): ngemva kokulanda i-repository ye-GGUF ehambisanayo, qala i-llama-server, bese usebenzisa i-python -m startlux_decision.gguf_server ukuze unikeze isevisi yokwenza izinqumo phambi kwe-llama.cpp.

Izinqumo ezisekelwe emathubeni: umbuzo ngamunye ubuyisa ngokuqondile amathuba ezinketho, ngakho ukuzethemba kuyatholakala ngokwemvelo ngaphandle kokulinganisa okwengeziwe.
Ukubambezeleka okuphansi kakhulu: impendulo itholakala nge-forward pass eyodwa; imodeli ye-4B ithatha ama-26ms kuphela, kanti i-35B-A3B ithatha ama-52.5ms kuphela.
Ukuhola kuwo wonke ama-benchmark: i-Decision Index 0.2.1 ifinyelela ku-63.88, idlule bonke abancintisana nabo abashicilelwe, kuhlanganise ne-Jev 1.13.
Umongo omude kakhulu: lonke uchungechunge lusekela ngokwemvelo okokufaka kwama-token angu-256K, futhi imibhalo emide ifundwa kanye kuphela.
Ubufakazi obunemodi eziningi: wonke amasayizi angafunda izithombe njengesisekelo sokwenza izinqumo.
Ukunqoba kuwo wonke amazinga afanayo: kusukela ku-0.8B kuya ku-35B, wonke amazinga asebenza kangcono kunamamodeli ancintisanayo anosayizi ofanayo.
Ukuhambisana ne-ecosystem ye-Jev: kusetshenziswa ifomethi ye-TypeSafe /v1/systemone, futhi amaklayenti abhalelwe i-Jev axhumeka ngaphandle kwanoma yiluphi ushintsho.

Indawo ye-GitHub: https://github.com/StartLuxLabs/Startlux-Decision
Indawo yamamodeli e-Hugging Face: https://huggingface.co/collections/startlux-models/startlux-decision-6abba92b301b573fa154d493

Ukwabela ngokuzenzakalelayo amathikithi kusevisi yamakhasimende: faka umbhalo noma isithombe sethikithi, bese ngesicelo esisodwa unqume ithimba elizosingatha, ukuthi kufanele kusheshiswe yini, kanye nobukhulu benkinga; amathuba ezinketho aba ukuzethemba.
Ukuzenzakalela kokusebenza kwekhompyutha: shayela isiphequluli sangempela ukuze ukhethe izilawuli ngesinyathelo ngesinyathelo futhi unqume ukuthi umsebenzi uqediwe yini; sekubonisiwe ukuqedela ngokuzenzakalelayo ukuthenga nokufaka ama-oda kuwebhu.
Izinqumo ze-Game AI: i-forward pass eyodwa ikhipha amathuba ezenzo ezifana nokuhambisa, ukwakha nokudubula; sekuqinisekisiwe ku-chess, i-StarCraft II, i-Doom nakweminye imidlalo.
Ukulawula ubungozi bezezimali nokuhambisana nemithetho: kuhlanganisa ukuhlola izigaba zezinkontileka, ukubona izinhlangano zezezimali, kanye nokuhlukanisa ama-imeyili obugebengu bokweba imininingwane; ukuzethemba kungasetshenziswa ngokuqondile ezinqubweni zokulawula ubungozi.
Ukusesha nokuhlukanisa kwamabhizinisi: ukuqaphela inhloso, ukuhlukanisa imibuzo, kanye nokuqinisekisa amaqiniso e-RAG kwenziwa ngokubambezeleka kwama-millisecond, okulungele izimo zebhizinisi ezinokucubungula okuningi ngesikhathi esisodwa.