- gpt-oss-20b ya zo azaman ƙirar nauyi mai buɗewa tare da aiwatar da gida da kuma dogon mahallin (har zuwa alamun 131.072).
- An inganta shi don NVIDIA RTX: An bayar da rahoton gudu zuwa 256 t / s; VRAM yana ɗauka don kula da aiki.
- Sauƙi don amfani tare da Ollama da madadin kamar su llama.cpp, GGML, da Microsoft AI Foundry Local.
- Hakanan ana samunsa a cikin filin wasa na Intel AI 2.6.0, tare da sabbin tsare-tsare da ingantaccen sarrafa muhalli.
A zuwa na gpt-oss-20b don amfanin gida yana kawo samfurin tunani mai ƙarfi wanda ke gudana kai tsaye akan PC ga ƙarin masu amfani. Wannan turawa, yayi daidai da Ingantawa don NVIDIA RTX GPUs, yana buɗe ƙofa zuwa buƙatun ayyukan aiki ba tare da dogaro da gajimare ba.
Mayar da hankali a bayyane yake: bayarwa bude-nauyi tare da mahalli mai tsayi sosai don hadaddun ayyuka kamar su ci-gaba bincike, bincike, code taimako ko dogon hira, fifita da sirri da sarrafa farashi lokacin aiki a gida.
Menene gpt-oss-20b ke bayarwa lokacin gudanar da gida?
Iyalin gpt-oss sun fara halarta tare da samfuran bude nauyi tsara don a sauƙaƙe haɗawa cikin hanyoyin magance ku. Musamman, gpt-oss-20b Ya fice don daidaita ƙarfin tunani da madaidaitan buƙatun kayan masarufi don PC na tebur.
Siffa mai bambanta ita ce shimfidar mahallin taga, tare da tallafi har zuwa 131.072 alamu a cikin kewayon gpt-oss. Wannan tsayin yana sauƙaƙe doguwar tattaunawa, Binciken takardu masu ƙarfi ko zurfafa zurfafa tunani ba tare da yanke ko rarrabuwa ba.
Idan aka kwatanta da rufaffiyar ƙira, buɗaɗɗen nauyi shawara yana ba da fifiko ga sassaucin haɗin kai a aikace: daga mataimaka tare da kayan aiki (wakilai) ko da plugins don bincike, binciken yanar gizo da shirye-shirye, duk suna cin moriyar fahimtar gida.
A cikin m sharuddan, kunshin na gpt-oss: 20b yana kusa da 13 GB shigar a cikin shahararrun wuraren runtime. Wannan yana saita sautin don albarkatun da ake buƙata kuma yana taimakawa wajen daidaitawa VRAM don kula da aiki ba tare da kwalabe ba.
Hakanan akwai babban bambance-bambancen (gpt-oss-120b), wanda aka tsara don yanayi tare da ƙarin wadatattun albarkatun hoto. Ga mafi yawan PC, duk da haka, da 20B Shi ne mafi mahimmancin farawa saboda dangantakarsa tsakanin sauri, ƙwaƙwalwa da inganci.
Ingantawa don RTX: Sauri, Ma'ana, da VRAM
Daidaita samfuran GPT-OSS zuwa yanayin muhalli NVIDIA RTX damar don high tsara rates. A cikin kayan aiki masu inganci, kololuwar har zuwa 256 tokens/second tare da gyare-gyare masu dacewa, yin amfani da ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun ƙayyadaddun abubuwa kamar su Farashin MXFP4.
Sakamako ya dogara da katin, mahallin, da daidaitawa. A cikin gwaje-gwaje tare da a RTX 5080, gpt-oss 20b ya kai kusa 128 t/s tare da abubuwan da ke ƙunshe (≈8k). Ta hanyar haɓaka 16k taga da kuma tilasta wasu kaya a cikin tsarin RAM, adadin ya ragu zuwa ~50,5 t/s, tare da GPU yin mafi yawan aikin.
Darasi a bayyane yake: da Dokokin VRAM. A cikin gida AI, a RTX 3090 tare da ƙarin ƙwaƙwalwar ajiya Zai iya yin aiki mafi kyau fiye da sabon GPU amma tare da ƙarancin VRAM, saboda yana hana ambaliya zuwa ga tsarin ƙwaƙwalwar ajiya da karin shiga tsakani na CPU.
Don gpt-oss-20b, ya dace don ɗaukar girman samfurin azaman tunani: game da 13 GB karin dakin ga KV cache da ayyuka masu tsanani. A matsayin jagora mai sauri, ana bada shawarar samun 16 GB na VRAM a kalla kuma ya nufa 24 GB idan ana sa ran abubuwa masu tsayi ko dorewa.
Masu neman matsi da kayan aikin zasu iya bincike ingantattun daidaito (kamar MXFP4), daidaita tsayin mahallin ko komawa zuwa saitunan GPU da yawa idan ya yiwu, koyaushe kiyaye burin kauce wa musanya zuwa RAM.
Shigarwa da amfani: Ollama da sauran hanyoyi
Don gwada samfurin a hanya mai sauƙi, Ollama yana ba da ƙwarewa kai tsaye akan PC masu ƙarfin RTX: Yana ba ku damar zazzagewa, gudanar da hira da GPT-OSS-20B ba tare da haɗaɗɗiyar jeri ba., ban da tallafawa PDFs, fayilolin rubutu, tsokanar hoto, da daidaita yanayin mahallin.
Hakanan akwai madadin hanyoyi don masu amfani da ci gaba, misali Sanya LLM akan Windows 11. Tsarin kamar kira.cpp da kuma buga dakunan karatu GGML an inganta su don RTX, tare da ƙoƙarin kwanan nan a rage nauyin CPU kuma ku ci riba CUDA Graphs. A cikin layi daya, Microsoft AI Foundry Local (a cikin samfoti) Haɗa samfura ta hanyar CLI, SDK ko APIs tare da haɓaka CUDA da TensorRT.
A cikin yanayin yanayin kayan aiki, Filin wasa na Intel AI 2.6.0 ya haɗa gpt-oss-20b a cikin zaɓuɓɓukan saSabuntawa yana ƙara ingantaccen sarrafa juzu'i don abubuwan baya da bita ga tsarin kamar BudeVINO, ComfyUI y kira.cpp (tare da goyon bayan aman wuta da daidaita mahallin), sauƙaƙewa kwanciyar hankali na gida.
A matsayin jagorar farawa, bincika Akwai VRAM, zazzage bambance-bambancen samfurin da ya dace da GPU ɗinku, tabbatar da alamar sauri tare da turawa wakilci da daidaitawa mahallin taga don kiyaye duk kaya akan katin zane.
Tare da waɗannan guda, yana yiwuwa a gina mataimaka don bincike da bincike, kayan aikin bincike ko goyon bayan shirin waɗanda ke gudana gaba ɗaya akan kwamfutar, suna kiyaye ikon mallakar bayanai.
Haɗin gpt-oss-20b tare da haɓaka RTX, kula da VRAM mai hankali, da kayan aiki kamar Ollama, llama.cpp, ko AI Playground cement wani zaɓi mai girma don gudanar da tunani AI a gida; hanyar da ke daidaita aiki, farashi, da keɓantawa ba tare da dogaro da sabis na waje ba.
Ni mai sha'awar fasaha ne wanda ya mayar da sha'awar "geek" zuwa sana'a. Na shafe fiye da shekaru 10 na rayuwata ta yin amfani da fasaha mai mahimmanci da kuma yin amfani da kowane nau'i na shirye-shirye saboda tsantsar son sani. Yanzu na kware a fasahar kwamfuta da wasannin bidiyo. Wannan shi ne saboda sama da shekaru 5 ina yin rubutu a gidajen yanar gizo daban-daban kan fasaha da wasannin bidiyo, ina ƙirƙirar kasidu da ke neman ba ku bayanan da kuke buƙata cikin yaren da kowa zai iya fahimta.
Idan kuna da wasu tambayoyi, ilimina ya bambanta daga duk wani abu da ya shafi tsarin aiki na Windows da kuma Android don wayoyin hannu. Kuma alƙawarina shine a gare ku, koyaushe a shirye nake in yi ƴan mintuna kaɗan in taimaka muku warware duk wata tambaya da kuke da ita a duniyar Intanet.