Ngokuhambisana ne-Claude Code: I-DeepSeek Harness isesivivinyweni somphakathi, kuvulwa kanye nohlelo lwama-plugin lwe-npm

I-DeepSeek-V4-Pro-0813 isivulwe ngokusemthethweni, kanti iphinde yakhipha inguqulo yokubuka yabathuthukisi yohlaka lwayo oluvulekile lwezisebenzeli zekhodi, i-Harness (esekelwe ku-Cordis), futhi yavula kanye nohlelo lwama-plugin oluhambisanayo.

Ngesikhathi i-IT之家 ibhala le ndaba, amakhosombe ahambisanayo e-GitHub ayesavaliwe emphakathini (isibuyekezo: manje asevuliwe). Abasebenzisi abanentshisekelo bangafaka leli thuluzi ngomyalo we-npm:

npm i @deepseek-ai/dsh

Ukuyiqalisa nge-npm:

npx @deepseek-ai/dsh web

Lo myalo uzoqalisa i-Web UI, ikheli elizenzakalelayo kube ngu-http://127.0.0.1:3080 .

Uma udinga ukuyiqalisa usebenzisa ikhodi yomthombo yekhosombe:

git clone https://github.com/deepseek-ai/deepseek-harness.gitcd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Incazelo elula ye-“Harness” iyithuluzi eliguqula imodeli ibe yisisebenzeli. Ngaphandle kwemodeli yolimi enkulu, i-Harness iphethe ukuhlela umongo, amathuluzi, isimo somsebenzi, impendulo nemingcele, ihlanganise amakhono emodeli endaweni yangempela futhi iqedele wonke umjikelezo kusukela ekuqondeni izidingo kuya ekulethweni kwekhodi. Phambilini i-DeepSeek yanikeza ifomula ethi “Model+Harness=Agent” olwazini lwayo olusemthethweni lokuqasha.

Ngokwendawo yomkhiqizo, i-DeepSeek Harness iqondiswe ngqo ku-OpenAI Codex kanye ne-Anthropic Claude Code, ihlose ukuba ithuluzi le-AI lokukhiqiza eligxile ezimweni zokuhlela nokusebenza ehhovisi.

Umholi weqembu u-Cui Tianyi wethule ngo-Agasti 2 isimemo esivulelekile emhlabeni wonke kubasebenzisi abazohlola i-Harness ngaphambi kokukhishwa. Ngemva kokushicilelwa kwesimemo, amashumi onjiniyela abavela emhlabeni wonke abhalisa ngesikhathi esifushane.

I-DeepSeek Harness (DSH) inikeza ithuluzi lomugqa womyalo elithi dsh, elisetshenziselwa ukumisa nokuqalisa umkhiqizo. Umyalo wayo oyinhloko, dsh --profile <name>, usetshenziselwa ukuqala ifayela lokumisa elikhethiwe. Leli thuluzi lisekela izindlela eziningi zokusebenza, ezihlanganisa i-Web ne-Headless, futhi livumela ukunwetshwa ngohlelo lwama-plugin.

I-DeepSeek-V4-Pro-0813 iyimodeli esekelwe kuma-Mixture of Experts (MoE), enenani lamapharamitha eliphelele elifinyelela ku-1.6 trillion. Ekucabangeni kwangempela, i-token ngayinye ivula amapharamitha ayizigidi eziyizinkulungwane ezingu-49 kuphela. Imodeli inewindi lomongo lama-token ayizigidi ezingu-1, futhi isekela ubude bokukhipha obufinyelela kuma-token ayizinkulungwane ezingu-384.

Ngokwendlela yokuxhumana, i-V4 Pro inikeza izinketho ezimbili, “imodi yokucabanga” kanye “nemodi engacabangi”, kanti imodi yokucabanga iyona ezenzakalelayo. Ngaphezu kwalokho, i-V4 Pro isekela ngokugcwele okukhiphayo okuhlelekile kwe-JSON, i-Tool Calls, i-Responses API kanye ne-Anthropic API neminye imisebenzi ethuthukile.

Ngokokuhlolwa kwangaphakathi okwenziwe yi-DeepSeek, inguqulo esemthethweni ithuthuke kakhulu uma iqhathaniswa nenguqulo yokubuka, futhi isondela noma idlule i-Claude Fable 5 ezivivinyweni eziningi. Isibonelo, esivivinyweni se-Terminal Bench 2.1 ithole u-87.9, yadlula kakhulu u-72.1 wenguqulo yokubuka. Kwezinye izivivinyo, i-DeepSWE inyuke isuka ku-12.8 yaya ku-62.7, kanti i-DSBench-Hard yaphindeka kabili yafinyelela ku-67.2.

Okubalulekile ukuthi izisindo ze-DeepSeek-V4-Pro-0813 sezivulwe ngokusemthethweni. Izisindo ze-DeepSeek zisebenzisa ilayisensi ye-MIT. Lokhu kuhunyushwe njengokuhlukanisa isisekelo semodeli ekuncikeni emandleni athile okwenza ikhompuyutha. Abathuthukisi bangayilanda bazakhele eyabo, bangabe besancika kusevisi eyodwa esefwini; ama-GPU akhiqizwa e-China namakhadi amandla ekhompyutha ajwayelekile angayisebenzisa ngokuqondile ngaphandle kokwenza ngokwezifiso komkhiqizi. Lokhu kwehlisa umgoqo wokusetshenziswa kwangasese emabhizinisini nakuhulumeni futhi kukhuphule amandla okumelana nezingozi zochungechunge lokuhlinzeka.