SenseNova Loop-9B-Preview Uyasebenza

I-SenseNova Loop-9B-Preview iyimodeli yokubuka kuqala ye-multimodal enezinga lamapharamitha angu-9B, ekhishwe yi-SenseTime njengomthombo ovulekile. I-SenseNova Loop-9B-Preview igcina izendlalelo zokudekhowuda zoqobo ezi-4 ekuqaleni nasekupheleni, kanti izendlalelo ezimaphakathi ziguqulwa zibe amamojuli okubala aphindaphindwayo abelana ngamapharamitha, ngokuphindaphinda imijikelezo eminingi ukuze yandise ukujula kokubala okusebenzayo, bese yethula indlela ye-mHC ukuze iqinise ukuzinza kokusebenzisana kwezici. Ngokwezivivinyo ezisemthethweni, imodeli idlula i-Qwen3.5-27B emisebenzini yasehhovisi neyama-agent, ifinyelela amakhono emisebenzi eyinkimbinkimbi asezingeni lamamodeli amakhulu kusetshenziswa amapharamitha ambalwa. Ama-weight ayo asevele eshicilelwe ku-GitHub naku-Hugging Face.

Ukukhiqiza ikhodi: Ifinyelela ezingeni elihamba phambili phakathi kwamamodeli alingana nayo emisebenzini yokubhala ikhodi, isekele izinhlelo zokusebenza zokuthuthukisa.
Ukukhiqiza okuqukethwe kwehhovisi: Uma ixhunywe kwiklayenti i-Xiaohuanxiong, ingakhiqiza izinto zokulethwa zasehhovisi ezifana nemiklamo ye-PPT nemibhalo emide.
Imisebenzi yama-agent: Ibonisa ukusebenza okuvelele ezivivinyweni zama-Agent ezinezinyathelo eziningi ezifana ne-AutomationBench, futhi ingaqedela imisebenzi emide eyinkimbinkimbi enokulandisa nokusebenzisana.
Ukucubungula umongo omude: Ngokwesikhundla esisemthethweni, iyimodeli ye-multimodal ephindaphindwayo eyenzelwe umongo omude nokugeleza komsebenzi kwama-agent.
Ukuhlola ukusetshenziswa komthombo ovulekile ngokwentengiso: Ama-weight nekhodi yokwenza inference ashicilelwe ku-GitHub naku-Hugging Face ukuze kusetshenziswe umphakathi ocwaningweni nasezinhlelweni zansuku zonke.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lezingxoxo lamaphrojekthi e-AI omthombo ovulekile
Isakhiwo se-Looped Transformer esiphindaphindwayo: Imodeli isekelwe esikalini se-9B, igcina amamojuli okuqala okudekhowuda angu-4 ekuqaleni nasekupheleni, bese ishintsha izendlalelo ezimaphakathi ibe amamojuli okubala aphindaphindwayo abelana ngamapharamitha. Iqoqo elifanayo lama-weight lisetshenziswa kabusha ngemijikelezo eminingi, ngaleyo ndlela kwandiswe ukujula kokubala okusebenzayo ngaphandle kokushintsha usayizi wamapharamitha, okwenza imodeli encane ithole ikhono lokucabanga elifana nenethiwekhi ejulile.
Indlela yokusebenzisana kwezici ye-mHC: Ukuze kuthuthukiswe ukudluliswa kolwazi phakathi kwemijikelezo eminingi, imodeli yethula i-Manifold-Constrained Hyper-Connections (mHC). Ihlanganisa izici kusetshenziswa imifudlana emibili yezinsalela ezinemizila eminingi kanye namamatiriksi okuxuba alawulwa umbandela we-“double stochastic”, iqinise ukusebenzisana kwezici phakathi kwezendlalelo futhi ilawule i-manifold yolwazi ukuze kugcinwe ukuzinza kokusakazeka kwesignali. Ukuhlolwa kokususa izingxenye kubonisa ukuthi le ndlela inyusa i-GPQA Diamond ngamaphuzu angu-4.42 ne-TeamBench ngamaphuzu angu-13.34.
Ukwabelana ngamapharamitha nokusebenza kahle: Ukwabelana ngama-weight ezendlalelo ezimaphakathi eziphindaphindwayo kwenza imodeli enama-paramethara acishe abe ngu-9B ithwale ukujula kokubala okudlula kakhulu imodeli eminyene engu-9B. Ngokuhambisana nobuchwepheshe obufana ne-TBK, kulawulwa izindleko zokubala zomjikelezo ngamunye, ngenhloso yokusondela ekusebenzeni kwemodeli esezingeni le-27B emisebenzini eyinkimbinkimbi ngenkathi kugcinwa izindleko ze-inference zilawuleka.

Ukulungiselela imvelo: Faka i-PyTorch ehambisana ne-CUDA, bese usebenzisa i-pip install "transformers==5.18.0" accelerate safetensors.
Ukuthola imodeli: Landa ama-weight e-sensenova/SenseNova-Loop-9B-Preview ku-Hugging Face uwagcine endaweni yangakini, uqinisekise ukuthi ifolda iqukethe ukucushwa, i-tokenizer, ithempulethi yengxoxo kanye nekhodi yemodeli yangokwezifiso.
Ukulayisha imodeli: Ekhodini, sebenzisa i-AutoModelForCausalLM.from_pretrained enama-trust_remote_code=True kanye ne-dtype=torch.bfloat16 ukuze ulayishe imodeli, bese ubiza i-.eval().
Ukwakha okokufaka: Sebenzisa i-tokenizer.apply_chat_template ukuhlanganisa i-prompt yengxoxo; uma kudingeka ukucabanga okujulile, setha i-enable_thinking=True.
Ukwenza i-inference: Biza i-model.generate ukuze ukhiqize umbhalo; i-use_cache iyasekelwa ukusheshisa, futhi i-MTP acceleration iyatholakala.
Ukuyifaka njengesevisi: Ezindaweni zokukhiqiza, i-LightLLM noma i-SGLang ingasetshenziswa ukunikeza i-API kubasebenzisi bangaphandle.

Ukusebenza kahle kwamapharamitha: Inamapharamitha acishe abe ngu-9B, futhi ngokwabelana ngama-weight nokuphindaphinda ingalingisa ukujula kokusebenza kwezendlalelo ezingu-56, isondele ekusebenzeni kwemodeli esezingeni le-27B ngezindleko eziphansi.
Ukuqamba kwesakhiwo: Isebenzisa i-Looped Transformer + i-mHC enemifudlana emibili, lapho ukuxutshwa kwezinsalela ezinemizila eminingi okulawulwayo kuthuthukisa ukusebenzisana kolwazi futhi kuqinise ukusakazeka kwesignali; ingenye yezindlela zokuqala ezisebenzisekayo zomthombo ovulekile ezisezingeni le-9B e-China.
Ikhono lomongo omude: Ukuqeqeshwa kusekela umongo omude ofika ku-262,144token, ofanele ukuqonda imibhalo nemisebenzi emide enezinyathelo eziningi.
Okokufaka kwe-multimodal: Isekela okokufaka kombhalo nezithombe, futhi ichazwa ngokusemthethweni njengemodeli ye-multimodal eyenzelwe umongo omude nokugeleza komsebenzi kwama-agent.
Ilungele ukusheshiswa kwe-inference: Inendlela ye-MTP acceleration eyakhelwe ngaphakathi futhi esivele isetshenziswa; izixazululo ze-speculative decoding ezifana ne-Dspark ne-Dflash2 zisacwaningwa.

Inqolobane ye-GitHub：https://github.com/OpenSenseNova/SenseNova-Loop
Inqolobane yemodeli ye-HuggingFace：https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview

Ukuqonda nokuhlaziya imibhalo emide: Isebenzisa umongo omude kakhulu ongu-256K, ingacubungula ngesikhathi esisodwa inkontileka yonke, umbiko wezezimali, iqoqo lamaphepha ocwaningo kanye neminye imibhalo emikhulu, bese ikhiqiza izifinyezo ezihlelekile nezimpendulo zemibuzo.
Ukugeleza komsebenzi kwama-agent: Ibonisa ukusebenza okuvelele ezimweni zokubiza amathuluzi ngezinyathelo eziningi, ukusebenza kumawebhusayithi nokuhlela imisebenzi, ngakho ifanele ukwakha ama-Agent okwenza imisebenzi yasehhovisi nezinqubo ngokuzenzakalela.
Ukukhiqiza okuqukethwe kwehhovisi: Uma ixhunywe kwiklayenti i-Xiaohuanxiong, ingakhiqiza ngokuqondile imiklamo ye-PPT, imibiko yokuhlaziya idatha namakhasi ezethulo anemiphumela yokunyakaza, ihlanganise ukukhiqizwa kwezinto zokulethwa zasehhovisi zansuku zonke.
Usizo ekuthuthukiseni ikhodi: Ikhono lokukhiqiza ikhodi lifinyelela ezingeni elihamba phambili phakathi kwamamodeli alingana nayo, futhi ingenza ukuqedela ikhodi, ukubhala ama-script nokuthuthukisa amaphrojekthi amancane kusukela ekuqaleni kuya ekupheleni.
Ukudala okuqukethwe kwe-multimodal: Isekela okokufaka kombhalo nezithombe, futhi ifanele ukukhiqiza okuqukethwe kwemidiya enothile okufana nezihloko zokuvakasha ezihlanganisa umbhalo nezithombe, amakhasi okuthengisa kanye namawebhusayithi alandisayo.