Xiaomi-OCR-0 Uyasebenza

I-Xiaomi-OCR-0 iyimodeli elula ye-multimodal document understanding engu-0.8B ekhishwe i-Xiaomi njengomthombo ovulekile, eqeqeshwe ku-Qwen3.5-0.8B, ehlanganisa ukuhlaziya amadokhumenti, ukukhishwa kolwazi olubalulekile, kanye ne-OCR visual question answering. Itholile amaphuzu angu-95.24 ku-Real5-OmniDocBench futhi yaba ngowokuqala, amaphuzu angu-96.83 ku-OmniDocBench v1.6 futhi yaba ngowesibili, kanti isilinganiso samaphuzu angu-83.2 emisebenzini emihlanu ye-OCR-VQA sidlule amamodeli asezingeni elifanayo. Injini yedatha ye-Xiaomi-OCR-0 isebenzisa ukulebula ngokuvumelana kochwepheshe kanye nokuqinisekisa ngokunikeza kabusha, futhi ihambisana namathuluzi e-MCP ne-Agent Skill. Seyisetshenziswa kumakhalekhukhwini be-Xiaomi ku-“Super XiaoAi”.

Ukuhlaziya amadokhumenti: Ihlonza iphinde ibuyisele umbhalo oyinhloko, amatafula ayinkimbinkimbi namafomula ezibalo, ikhiphe imiphumela ehlelekile ye-Markdown/LaTeX ngokulandelana kwemvelo kokufunda.
Ukukhishwa kolwazi olubalulekile (KIE): Kumadokhumenti anesakhiwo esiyingxenye, njengama-invoyisi namakhadi omazisi, ithola ngokunembile izinkambu ezicacisiwe futhi ikhiphe idatha ehlelekile ngefomethi ye-JSON.
I-OCR visual question answering: Yeqa ukudlulisa umbhalo kuphela, isekele ukusesha phakathi kwezakhi, ukubala amanani, kanye nemibuzo nezimpendulo zokucabanga ngencazelo kumadokhumenti namashadi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxoxa ngamaphrojekthi e-AI omthombo ovulekile
Ukulebula ngokuvumelana kochwepheshe abahlukene (ETC): Amamodeli angu-8 e-OCR anokwakheka, idatha namasu ahlukene abikezela ngokuzimela izindawo zamadokhumenti. Ilinganisa ukuvumelana ngababili kombhalo (NED), amatafula (TEDS) namafomula (CDM), ikhethe ama-triplet amathathu anokuvumelana okuphezulu bese isebenzisa imodi ephakathi kwawo njengamalebula-mbumbulu; amasampula anokuzethemba okuphezulu kuphela, anokuvumelana okungu-≥0.9, angena ngqo kusethi yokuqeqesha, ngaleyo ndlela kuhluzwe ukuchema kwemodeli eyodwa kusukela ekuqaleni.
Ukuhlola okuholwa ukunikezwa kabusha (Render-Guided Refine-and-Judge): Ezindaweni lapho imibono yochwepheshe ingavumelani khona, i-Qwen3.5-122B-A10B inikeza kabusha imiphumela yokuhlaziya engabangwa ibe izithombe, bese iyiqhathanisa nendawo yokuqala ezingeni lamaphikseli nesakhiwo ukuze ithole amaphutha acashile njengokungahleleki kwemigqa namakholomu noma ukushoda kwezimpawu. Ilungisa ngokuphindaphindiwe kuze kufike emijikelezweni emi-3; okungaguquki kudluliselwa esigabeni sesibili esinomlando ophelele wokuqinisekisa, lapho kudlula khona “ukuxilonga—ukulungisa—ukuqinisekisa”. Ekugcineni kumele kubuyele echibini lochwepheshe ukuze kufinyelelwe ekuvumelaneni ngaphambi kokwamukelwa.
Ukumba amasampula nokuhlanganisa okuqondisiwe: Isilinganiso sokuqeqesha sinqunywa ngokuhlanganyela yizimpawu ezintathu: ukuvumelana kochwepheshe, ukuzivumelanisa kwemodeli nayo, kanye nokuhlanganiswa kwencazelo nge-K-Means. Ngokwalokho kwakhiwa isethi exubile ehlanganisa kabanzi ye-CPT kanye nesethi yezibonelo ezinzima ye-RL; idatha ehlanganisiwe ihamba ngezindlela ezimbili, “eqhutshwa ukumbozwa” kanye “eqhutshwa ukwehluleka”, futhi ukwakheka kwamukelwa kuphela ngemva kokuqinisekiswa kusethi yokuqinisekisa emisiwe.
Ukuqeqeshwa kwezigaba ezintathu okuqhubekayo: Okokuqala kusetshenziswa i-Q-Mask text anchoring pretraining ukusungula ukuqondanisa okunemininingwane phakathi kokuqukethwe kombhalo nendawo; bese kwenziwa i-multi-task OCR continual pretraining ukuze kwandiswe imisebenzi efana nokuhlaziya ikhasi lonke, ukuhlonza izindawo ne-VQA ngaphansi kwefomethi efanayo ye-autoregressive; ekugcineni kusetshenziswa i-DAPO-style Mix-RL reinforcement learning ezibonelweni ezinzima ezimbiwe, imivuzo ihanjiswa ngokohlobo lomsebenzi ukuze kudlule imingcele yezakhiwo eziyinkimbinkimbi nokucabanga okuyizinyathelo eziningi ngemivuzo eqinisekisekayo.

Thola ikhodi: Clone-a i-GitHub repository SeerRay-Lab/Xiaomi-OCR-0 bese ufaka indawo yokuncika echazwe ku-requirements.
Landa izisindo: Landa amafayela ezisindo zemodeli engu-0.8B ku-Hugging Face SeerRay-Lab/Xiaomi-OCR-0.
Ukuhlola ngokushesha: Sebenzisa iskripthi esihambisana nephrojekthi ukuze uqale i-Demo esebenzisanayo esipheqululini, bese ulayisha izithombe zamadokhumenti ukuze uzame ukuhlaziya, ukukhishwa kolwazi nokuphendula imibuzo ku-inthanethi.
Ukulungisa ukubizwa yi-Agent: Xhuma isevisi ye-MCP noma i-Agent Skill ekurepository ku-agent yakho ehlakaniphile ukuze wenze ngokuzenzakalelayo inqubo ethi “hlonza isikrini/umbhalo—khipha izinkambu—yenza isinyathelo esilandelayo”.
Ukusebenzisa isevisi yasendaweni: Sebenzisa i-SGLang noma i-vLLM ukuze usebenzise imodeli njengesevisi yasendaweni ye-inference, izinhlelo zebhizinisi ziyibize nge-API.
Ukuhlanganisa ukutholwa kwesakhiwo: Lapho usebenzisa ipayiphu yezigaba ezimbili, qala usebenzise i-PP-DocLayoutV3 ukuthola izindawo zombhalo, amatafula namafomula, bese uthumela indawo ngayinye ku-Xiaomi-OCR-0 ukuze ihlonzwe futhi ihlanganiswe ibe umphumela wekhasi lonke.
Ukubhala imiyalelo: Yakha izimfuno ngefomethi efanayo ye-image–instruction–response, ucacise ngokwehlukana uhlobo lomsebenzi: ukuhlaziya, KIE noma VQA.

Ukusebenza kahle okuphezulu kwamapharamitha: Ngamapharamitha angu-0.8B kuphela ithole amaphuzu angu-96.83 ku-OmniDocBench v1.6, futhi amamodeli akhethekile angu-0.8B–1.2B adlule ngokuphelele amamodeli amakhulu ajwayelekile afana ne-GPT-5.2 ne-Qwen3-VL-235B.
Ukumelana nezimo zangempela kuhamba phambili: Ithole amaphuzu angu-95.24 ku-Real5-OmniDocBench, yahola kuzo zonke izimo ezinhlanu zokuphazamiseka: ukuskena, ukugobeka kwephepha, ukuthwebula isikrini, ukukhanya okungalingani, nokuthwebula ngokutsheka.
Amakhono amathathu ahlanganisiwe: Ukuhlaziya amadokhumenti, ukukhishwa kolwazi olubalulekile ne-OCR visual question answering kuhlanganiswe kumodeli eyodwa; imisebenzi ishintshwa ngesixhumi esibonakalayo semiyalelo esabiwe, ngaphandle kokuhlanganisa amamodeli amaningi akhethekile.
Injini yedatha ithatha indawo yokunqwabelanisa amapharamitha: Amasampula ayizigidi ezingu-170 akhiqizwa ngokuzenzakalela ngokuvota kochwepheshe, ukuqinisekisa okuholwa ukunikezwa kabusha kanye nokuhlanganisa okuqhutshwa ukwehluleka, kusetshenziswa ikhwalithi yedatha esikhundleni sobukhulu bemodeli.
Ukufunda kwekharikhulamu nokudlulisa okuhle: Ngemva kokuvuthwa kokuhlaziya, ukuqondwa kolwazi kubuyisela inzuzo ekuhlaziyeni kuze kufike ku-+0.64 amaphuzu, okwenza kube nokuthuthuka okubili kwamakhono kumodeli elula.

Iwebhusayithi yephrojekthi：https://huggingface.co/spaces/xiaomi-research/Xiaomi-OCR-0
i-GitHub repository：https://github.com/xiaomi-research/Xiaomi-OCR-0
Ilabhulali yamamodeli ye-HuggingFace：https://huggingface.co/xiaomi-research/Xiaomi-OCR-0
Iphepha lobuchwepheshe le-arXiv：https://arxiv.org/pdf/2609.36136

Ukuphathwa kwamathikithi olwazi lwansuku zonke ngesikhathi esifanele: Ihlonza amakhodi okuthatha ukudla okulethiwe, amakhodi okuthatha amaphasela, izinombolo zolayini nolunye ulwazi olunesikhathi, bese “ilufaka esiqhingini” ukuze lufinyeleleke ngokuthinta kanye lapho ludingeka.
Ukubalwa kwemali okuhlakaniphile nokuhlelwa kwezikweletu: Ngokuchofoza kanye ikhipha inani, umthengisi, isikhathi nesigaba emininingwaneni yezikweletu ze-WeChat/Alipay, bese ikhiqiza ngokuzenzakalela amarekhodi okubalwa kwemali.
Ukufakwa okuhlelekile kwamakhadi nezitifiketi: Ngokuthatha isithombe noma ukubona isikrini, ikhipha izinkambu zikamazisi, ikhadi lasebhange, i-invoyisi, ipasipoti nezinye zibe yi-JSON, ithathe indawo yokufakwa nokuhlolwa ngesandla.
Ukwakhiwa kwesheduli okuhlakaniphile: Ikhipha isikhathi nendawo ezazisweni zezingxoxo zamaqembu nasemaposta emicimbi, ikhiqize ngokuzenzakalela ikhadi lesheduli bese ikwengeza kukhalenda.
Ukuxhumanisa ulwazi lohambo: Ikhipha imininingwane yohambo efana nenombolo yendiza ezingxoxweni, ihlole ngokuzenzakalela izibuyekezo zesikhathi sangempela zokufika nokusuka, bese ithumela izikhumbuzi.