MiniCPM5-2B Uyasebenza

I-MiniCPM5-2B iyimodeli eyisisekelo yolimi esebenza emadivayisini, ekhishwe ngomthombo ovulekile ngokubambisana yi-Mianbi Intelligent, umphakathi we-OpenBMB kanye ne-Tsinghua University. Imodeli inamapharamitha acishe abe yizigidi eziyinkulungwane ezingu-2.5, isekela umongo ongu-128K, futhi ithole isilinganiso esingu-53.9 kumabhentshimakhi angu-34, yadlula amamodeli amaningi ezinga le-4B futhi yafinyelela ku-SOTA yamamodeli omthombo ovulekile ezinga le-2B. Imodeli ifaka amakhono e-Agent afana nokubiza amathuluzi, ukusesha okujulile nokukhiqiza ikhodi emadivayisini, futhi ingasebenza kudivayisi enememori engu-6GB; ngesikhathi esifanayo ivula idatha yokuqeqesha, uhlaka lwe-RL i-Meshy kanye ne-algorithm ye-JustRL II ngomthombo ovulekile, okuyenza ibe “ithuluzi elincane elinamandla” emadivayisini.

Izingxoxo nezimpendulo emadivayisini: Qalisa imodeli yolimi ngqo kudivayisi yasendaweni ukuze wenze imisebenzi efana nezimpendulo zansuku zonke, ukubhala nokufingqa, ngaphandle kokulayisha idatha efwini.
Ukuqonda imibhalo emide: Isekela umongo ongu-128K, futhi ingacubungula endaweni amanothi emihlangano, izinkontileka nemibhalo emide, ikhiphe ulwazi olubalulekile ngokunembile.
Ukukhiqiza nokulungisa ikhodi: Kusukela ekuqedeni ikhodi kuya ekulungiseni okusezingeni lenqolobane yangempela (SWE-bench Verified 46.4), ingasebenza njengomsizi wokubhala ikhodi wendawo.
Ukubiza amathuluzi: Isekela ngokwendabuko ukubizwa kwemisebenzi, futhi ingabiza amathuluzi angaphandle lapho kudingeka ukuze yenze imibuzo, izibalo nokuxhumana nesizindalwazi.
Ukusesha okujulile: Inamandla e-Search Agent, ingahlela ngokuzimela ukusesha okuyizinyathelo eziningi futhi ihlanganise izimpendulo.
Ukucabanga okuxubile: Ngeswishi ye-enable_thinking, ungashintsha phakathi kwemodi yezimpendulo ezisheshayo nemodi yokucabanga okujulile, uhlanganise ukusebenza kahle nokuxazulula izinkinga ezinzima.
Imisebenzi yenqwaba yasendaweni: Imisebenzi ebucayi ngobumfihlo efana nokuhlunga ama-CV, ukumaka amaphepha nokuhlaziya izinkontileka ingacutshungulwa ngokuhambisana ngentengo ephansi ohlangothini lwedivayisi.

Landela i-WeChat bese uphendula ngo-“开源” ukuze ujoyine iqembu lezingxoxo zamaphrojekthi e-AI omthombo ovulekile
Ukwakhiwa okujwayelekile kwe-Dense Transformer: Imodeli isebenzisa ukwakheka okujwayelekile kwe-LlamaForCausalLM, inezingqimba ezingu-42, futhi isebenzisa i-GQA ukunciphisa izindleko ze-KV Cache. Isekela ngokwendabuko umongo wamathokheni angu-131,072. Ngenxa yokuthi ukwakheka kujwayelekile ngokuphelele, ama-framework adumile afana ne-vLLM, llama.cpp ne-Ollama angayilayisha ngqo ngaphandle kwe-kernel eyenziwe ngokwezifiso.
Ukuthuthukiswa kwamakhono e-Agent kuwo wonke umzila: I-Agent ilungiselelwa kusukela esigabeni sokuqeqeshwa kwangaphambili. Kuqala, isikali esingu-200B se-Agent Midtraining sifaka ulwazi lokuqala lokuhlukanisa imisebenzi nokusebenzisa amathuluzi; bese imikhondo ye-SFT engu-500,000 ehlanganisa ukubiza amathuluzi, ukusesha iwebhu ne-Code Agent ifundisa imodeli ukwenza imisebenzi ngendlela efanele; ekugcineni, i-Agent RL iqinisekisa ukuzinza kwamasu ekusebenzelaneni kwemijikelezo eminingi, kwakhiwe uchungechunge oluphelele lwamakhono “olusebenzayo nolusebenziseka kahle”.
Uhlaka lwe-reinforcement learning i-Meshy: I-Meshy isusa isilawuli esimaphakathi nokuncika ku-Ray ekuqeqeshweni kwe-RL okuvamile, futhi imodela ukuqeqesha, ukucabanga nokubalwa kwemiklomelo njengezinsizakalo zama-node alinganayo. Isebenzisa isimo sokulungela kwedatha ku-TransferQueue ukuqhuba ukugeleza kolawulo kwangempela, futhi ingashintsha ngokuguquguqukayo phakathi kwamamodi amathathu: synchronous, asynchronous kanye ne-fully asynchronous. Lokhu kwenza ukuqeqeshwa kwe-RL ngezinga elikhulu kube lula ukukwandisa futhi kube lula ukuyisabalalisa.
I-algorithm yokufunda ngemiklomelo i-JustRL II: Ibhekana nezinkinga ezimbili ezinkulu ku-RL yochungechunge olude lokucabanga yamamodeli asemaphethelweni: “isignali yomvuzo idukisa imodeli” kanye “nokwabiwa kwekhredithi kwe-GRPO akucacile”. I-JustRL II isebenzisa iphayiphi enezigaba ezintathu ohlangothini lwedatha ukulinganisa amasampula okuqeqesha, futhi yethula i-Critic ku-GRPO ohlangothini lwe-algorithm ukuze yabele ikhredithi ezingeni lamathokheni, ihlukanise ngokunembile umnikelo wesinyathelo ngasinye ochungechungeni olude lokucabanga.
Ukuhlanzwa kwedatha okusekelwe ekubizweni kwemisebenzi kwe-UltraX: I-UltraX iyayeka indlela evamile yokubhala kabusha umbhalo wewebhu ekupheleni kuya ekupheleni ngemodeli enkulu. Kunalokho, iqeqesha imodeli encane engu-0.6B ukubikezela imisebenzi yokuhlela ehlelekile, bese umenzi onqumayo eyisebenzisa umugqa ngomugqa. Lokhu kugwema ukushintsha kwencazelo nokonakala kwesakhiwo okubangelwa ukubhala kabusha, futhi ngenxa yokuthi imisebenzi yokuhlela ingagcinwa futhi ihlolwe, yonke inqubo iyalandeleka. Ekuhlolweni, umphumela wokuqeqesha ngedatha ehlanzwe yamathokheni angu-16B wedlula ukusebenza kokuqeqeshwa okuphelele kwedatha yokuqala yamathokheni angu-20B.
Uhlelo lokuphathwa kwedatha olunezigaba lwe-UltraData: Idatha yekhodi iphathwa ngokulandelana kwe-L0–L3. Idatha ye-RL idlula kuphayiphi enezigaba ezintathu esusa amasampula angenakuqinisekiswa, anemiklomelo engathembekile noma anobunzima obungalingani. Isignali yedatha esezingeni eliphezulu ivala ubuthakathaka bomthamo wemodeli encane, okuyisisekelo esibalulekile sobunjiniyela sokudlula amamodeli e-4B ngenani elincane lamapharamitha.

Qinisekisa izimfuneko zehadiwe: Isisindo sokuqala se-BF16 sicishe sibe ngu-4.7GiB, noma ungasebenzisa inguqulo ye-GGUF elinganisiwe engaba ngu-1.2GB kumakhalekhukhwini noma kwikhompyutha ephathekayo evamile enememori engu-6GB.
Thola izisindo zemodeli: Landa inguqulo oyidingayo ku-HuggingFace noma emphakathini we-Magic-Maker, bese ukhetha ifomethi ye-BF16 / GGUF / MLX / GPTQ / SFT njengoba kudingeka.
Faka i-framework yokucabanga: Ezimeni zefu noma zokuxhumana okuningi sebenzisa i-vLLM (isekelwa ngokwendabuko kusukela ku-≥0.21.0, futhi i-TP=1 yekhadi elilodwa yanele); ekusatshalalisweni kwasendaweni okulula sebenzisa i-llama.cpp, i-Ollama noma i-SGLang; endaweni ye-CPU ungasebenzisa i-framework i-Arclight eyakhiwe yi-Mianbi.
Layisha bese usebenzisa imodeli: Layisha ngqo usebenzisa ukwakheka okujwayelekile kwe-Llama, ngaphandle kwe-kernel eyenziwe ngokwezifiso.
Shintsha imodi yokucabanga: Sebenzisa iswishi ye-enable_thinking kuthempulethi yengxoxo ukulawula izimpendulo ezisheshayo noma ukucabanga okujulile, ukhethe imodi ngokuya ngobunzima bomsebenzi.
Xhuma ukubizwa kwamathuluzi/Agent: Sebenzisa ukubizwa kwemisebenzi nokusesha okujulile okwakhelwe kwimodeli, uhlanganise i-Cookbook yokusatshalaliswa ehambisanayo nama-Agent Skills ukuze wakhe ukugeleza komsebenzi we-Agent wendawo.

I-SOTA yomthombo ovulekile ezinga le-2B: Isilinganiso samabhentshimakhi angu-34 singu-53.9, idlula wonke amamodeli ezinga le-4B ahlolwe, ikhombisa ukuthi okuncane kunganqoba okukhulu.
Isisekelo se-Agent ejwayelekile emadivayisini: Ihlanganisa ngokwendabuko ukubiza amathuluzi, ukusesha okujulile nokukhiqiza ikhodi, phakathi kwamanye amakhono abalulekile e-agent. Iphambili ohlwini lwe-Agent ngamaphuzu angu-20, idlula amamodeli amakhulu afana ne-Granite 4.2 8B ne-Qwen3.5 9B.
Ukusatshalaliswa okulula kakhulu: Ukwakheka okujwayelekile kwe-Llama kanye nokukhishwa kwawo wonke amafomethi kwenza i-vLLM, i-Ollama ne-llama.cpp ilayishe ngqo ngaphandle kokuguqulwa. Inguqulo ye-INT4 ingu-1.2GB kuphela futhi isebenza kudivayisi enememori engu-6GB.
Umongo omude wendabuko ongu-128K: Umongo wamathokheni angu-130,000 unikeza isisekelo sokucubungula endaweni imibhalo emide efana namanothi emihlangano nezinkontileka.
Umthombo ovulekile kuwo wonke umzila: Akukhishwanga izisindo kuphela, kodwa nedatha yokuqeqesha (uchungechunge lwe-UltraData), uhlaka lwe-RL (Meshy), i-algorithm ye-RL (JustRL II) kanye ne-Recipe yokuqeqesha, okuvumela ukuphindaphindwa okuphelele nokusungula okusha.
Ukucabanga okusebenzayo: Umsebenzi owodwa usebenzisa ngokwesilinganiso amathokheni okukhipha angaba ngu-21K, kanti ukusetshenziswa kwamathokheni kuphansi kunamamodeli ezinga elifanayo, ngakho isebenza ngokushesha futhi yonga amandla emadivayisini.
Imvelo ebanzi yama-chip: Ihambisana kusukela ku-Day0 nama-chip asekhaya angu-9, ahlanganisa i-AMD, Intel, MediaTek, Qualcomm, Ascend, Hygon, Kunlunxin kanye ne-Moore Threads.

Inqolobane ye-GitHub：https://github.com/OpenBMB/MiniCPM
Inqolobane yamamodeli e-HuggingFace：https://huggingface.co/collections/openbmb/minicpm5

Ukucubungula imibhalo ebucayi ngobumfihlo: Imibuzo nezifinyezo zamanothi emihlangano, izinkontileka nama-imeyili angaphakathi kwenziwa endaweni, ngaphandle kokuthi idatha ishiye idivayisi.
Ukuqasha nezabasebenzi: Hlunga ama-CV amaningi endaweni, khipha ulwazi lwabafakizicelo bese uluqhathanisa ngendlela ehlelekile, ngaphandle kokulayisha ama-CV efwini.
Umsizi wokubhala ikhodi ohlangothini lwedivayisi: Qedela ikhodi, unikeze iziphakamiso zokulungisa amaphutha futhi ulungise amaphutha ezingeni lenqolobane ungaxhunyiwe ku-inthanethi, uvikele impahla yekhodi.
Ukusesha okujulile kwasendaweni: Hlela ngokuzenzakalelayo ukusesha kwewebhu okuyizinyathelo eziningi bese uhlanganisa izimpendulo, okulungele ucwaningo, ukuqhathanisa amanani nokuqoqwa kolwazi.
Ezemfundo nokumaka: Maka amaphepha, nikeza impendulo ngomsebenzi wesikole futhi uhlaziye amaphutha ngesikhathi esisodwa endaweni, wehlise izindleko zezikole nezikhungo zokuqeqesha.