Mafi Kyawun Buɗaɗɗen Tushen Samfuran Samar da Rubutu

Jagora mai amfani, wanda ya fi fifita masu gini wajen zaɓar mafi kyawun samfuran samar da rubutu kyauta—tare da bayyanannun fa'idodi da rashin fa'idodi, zaɓuɓɓuka masu sauri bisa yanayi, da hanyoyi masu sauƙi don gwada su a cikin ShareAI Playground.

---

TL;DR

Idan kana son mafi kyawun samfuran samar da rubutu na buɗaɗɗen tushe a yanzu, fara da fitowar da aka daidaita don umarni mai sauƙi don juyawa da sauri da ƙarancin farashi, sannan ka haɓaka kawai idan an buƙata. Ga mafi yawan ƙungiyoyi:

• Gwaji mai sauri (na'ura mai kwakwalwa/laptop/CPU mai dacewa): gwada samfuran da aka daidaita don umarni masu nauyi 1–7B; rage su zuwa INT4/INT8.
• Ingancin matakin samarwa (daidaitaccen farashi/jinkiri): samfuran tattaunawa na zamani 7–14B tare da tsawon mahallin da ingantaccen ajiyar KV.
• Yawan aiki a matakin girma: haɗin-gwanaye (MoE) ko samfuran masu yawa masu inganci a bayan ƙarshen da aka shirya.
• Yare da yawa: zaɓi dangin da ke da ƙarfi wajen horar da rubutu ba na Ingilishi ba da haɗin umarni.

👉 Bincika samfuran sama da 150 a kan Kasuwar Samfura (matattara don farashi, jinkiri, da nau'in mai bayarwa): Duba Samfura

Ko ka tsallake kai tsaye zuwa cikin Filin wasa ba tare da kayan aiki ba: Gwada a Filin Gwaji

Ka'idodin Kimantawa (Yadda Muka Zaɓa)

Alamomin ingancin samfurin

Muna neman bin umarni mai ƙarfi, samar da tsawon lokaci mai ma'ana, da alamomin gwaji masu gasa (tunanin hankali, lamba, taƙaitawa). Kimantawar ɗan adam da ainihin tambayoyi sun fi mahimmanci fiye da hotunan jagorar.

Bayyanar lasisi

“Buɗe tushen” ≠ “buɗaɗɗen nauyi.” Muna son lasisi mai sassauci irin OSI don amfani da kasuwanci, kuma muna bayyana a sarari lokacin da samfurin ke da nauyi-buɗe kawai ko yana da takunkumin amfani.

Bukatun kayan aiki

Kasafin kuɗin VRAM/CPU yana tantance abin da “kyauta” ke nufin a zahiri. Muna la'akari da samuwar quantization (INT8/INT4), girman taga mahallin, da ingancin KV-cache.

Balagar yanayi

Kayan aiki (sabar samarwa, tokenizers, adapters), goyon bayan LoRA/QLoRA, samfuran tambaya, da kulawa mai aiki duk suna shafar lokacin zuwa darajarku.

Shirye-shiryen samarwa

Jinkirin ƙarshen ƙarami, kyawawan tsoffin tsaro, lura (token/jinkirin ma'auni), da daidaiton hali a ƙarƙashin nauyi suna yin ko karya ƙaddamarwa.

Manyan Buɗaɗɗen Samfuran Samar da Rubutu (Kyauta don Amfani)

Kowanne zaɓi da ke ƙasa ya haɗa da ƙarfi, yanayin amfani da ya dace, bayanan mahallin, da shawarwari masu amfani don gudanar da shi a gida ko ta hanyar ShareAI.

Iyali Llama (buɗaɗɗen bambance-bambance)

Dalilin da ya sa yake nan: An karɓa sosai, halayen hira mai ƙarfi a cikin ƙananan zuwa matsakaicin kewayon sigogi, wuraren bincike masu daidaitawa da umarni, da babbar tsarin haɗin gwiwa na kayan aiki da kayan aiki.

Mafi dacewa ga: Hira ta gabaɗaya, taƙaitawa, rarrabuwa, tambayar kayan aiki (fitattun sakamako).

Mahalli & kayan aiki: Yawancin bambance-bambance suna tallafawa tsawaita mahalli (≥8k). Quantizations na INT4 suna gudana akan GPUs na yau da kullun na masu amfani har ma da CPUs na zamani don ci gaba/gwaji.

Gwada shi: Tace samfuran iyali Llama akan Kasuwar Samfura ko buɗe a cikin Filin wasa.

Jerin Mistral / Mixtral

Dalilin da ya sa yake nan: Tsarin ingantattu tare da bambance-bambancen hira masu daidaitawa da umarni; MoE (misali, salon Mixtral) yana ba da kyakkyawan ma'auni tsakanin inganci/latency.

Mafi dacewa ga: Hira mai sauri, mai inganci; taimako mai juyawa da yawa; daidaitawa mai tsada.

Mahalli & kayan aiki: Mai sauƙi ga quantization; bambance-bambancen MoE suna haskakawa lokacin da aka yi hidima yadda ya kamata (router + batching).

Gwada shi: Kwatanta masu samarwa da latency akan Duba Samfura.

Iyali Qwen

Dalilin da ya sa yake nan: Babban rufin harsuna da bin umarni; sabuntawa na al'umma akai-akai; aiki mai kyau na lamba/chat a cikin ƙananan girma.

Mafi dacewa ga: Tattaunawa da ƙirƙirar abun ciki a harsuna da yawa; tsararrun tambayoyi masu nauyi akan umarni.

Mahalli & kayan aiki: Kyawawan zaɓuɓɓukan ƙananan samfura don CPU/GPU; akwai nau'ikan dogon mahallin.

Gwada shi: Kaddamar da sauri a cikin Filin wasa.

Iyali Gemma (nau'ikan OSS masu sassauci)

Dalilin da ya sa yake nan: Halin da aka daidaita da umarni mai tsabta a cikin ƙananan sawun; abokantaka ga matukan jirgi akan na'ura; takardun ƙarfi da samfuran tambaya.

Mafi dacewa ga: Mataimaka masu nauyi, ƙananan hanyoyin samfur (cikakken rubutu, taimako na ciki), taƙaitawa.

Mahalli & kayan aiki: Ana ba da shawarar ƙididdige INT4/INT8 don kwamfutocin tafi-da-gidanka; lura da iyakokin alamar don ayyuka masu tsawo.

Gwada shi: Duba waɗanne masu samarwa ke ɗaukar nau'ikan Gemma akan Duba Samfura.

Iyali Phi (mai nauyi/mai rahusa)

Dalilin da ya sa yake nan: Samfura masu matuƙar ƙanƙanta waɗanda ke yin aiki fiye da girman su akan ayyukan yau da kullum; sun dace lokacin da farashi da jinkiri suka mamaye.

Mafi dacewa ga: Na'urorin gefen, sabobin CPU kawai, ko samarwa na rukunin layi.

Mahalli & kayan aiki: Yana son ƙididdigewa; mai kyau don gwaje-gwajen CI da duba hayaki kafin ka faɗaɗa.

Gwada shi: Gudanar da kwatancen sauri a cikin Filin wasa.

Sauran zaɓuɓɓuka masu ƙanƙanta masu ban sha'awa

• Samfuran tattaunawa masu daidaitawa na umarni 3–7B an inganta su don sabobin RAM masu ƙanƙanta.
• Samfuran dogon mahallin (≥32k) don tambayoyi na takardu da bayanan taro.
• Ƙananan samfuran masu karkata zuwa lambar kwamfuta don taimakon ci gaba na inline lokacin da manyan LLMs na lambar kwamfuta suka yi yawa.

Shawara: Don gudu akan kwamfutar tafi-da-gidanka/CPU, fara da INT4; haɓaka zuwa INT8/BF16 kawai idan inganci ya ragu don tambayoyinku.

Mafi kyawun Zaɓuɓɓukan “Matakin Kyauta” Masu Gudanarwa (Lokacin da Ba Ku So Ku Gudanar da Kanku ba)

Ƙarshen matakin kyauta suna da kyau don tabbatar da tambayoyi da UX, amma iyakokin ƙima da manufofin amfani da adalci suna fara aiki da sauri. Yi la'akari da:

• Ƙarshen al'umma/Masu bayarwa: ƙarfin lokaci-lokaci, iyakokin ƙima masu canzawa, da farawa masu sanyi lokaci-lokaci.
• Fa'idodi da rashin fa'ida idan aka kwatanta da na gida: masu gudanarwa suna cin nasara akan sauƙi da sikeli; na gida suna cin nasara akan sirri, jinkirin lokaci mai ƙayyadaddun lokaci (da zarar an dumama), da tsadar API mara ƙari.

Yadda ShareAI ke taimakawa: Hanyar zuwa masu bayarwa da yawa tare da maɓalli guda, kwatanta jinkiri da farashi, da canza samfura ba tare da sake rubuta aikace-aikacenku ba.

• Ƙirƙiri maɓallinka cikin dannawa biyu: Ƙirƙiri Maɓallin API
• Bi API mai sauri: Bayanin API

Teburin Kwatanta Mai Sauri

Iyali na samfur	Salon lasisi	Params (na yau da kullum)	Taga mahallin	Salon fassara	VRAM na yau da kullum (INT4→BF16)	Ƙarfi	Ayyuka masu dacewa
Iyali na Llama	Nauyin buɗe / nau'ikan izini	7–13B	8k–32k	GPU/CPU	~6–26GB	Tattaunawa ta gaba ɗaya, umarni	Mataimaka, taƙaitawa
Mistral/Mixtral	Nauyin buɗe / nau'ikan izini	7B / MoE	8k–32k	GPU (CPU ci gaba)	~6–30GB*	Daidaiton inganci/jinkiri	Mataimakan samfur
Qwen	OSS mai izini	7–14B	8k–32k	GPU/CPU	~6–28GB	Yare da yawa, umarni	Abun ciki na duniya
Gemma	OSS mai izini	2–9B	4k–8k+	GPU/CPU	~3–18GB	Ƙaramin, tsaftataccen hira	Matukan kan-na'ura
Phi	OSS mai izini	2–4B	4k–8k	CPU/GPU	~2–10GB	Ƙanana & masu inganci	Gefe, ayyukan tsari

Yadda za a Zaɓi Mafi Kyawun Samfuri (Yanayi 3)

1) Fara jigilar MVP akan kasafin kudi

• Fara da ƙaramin samfurin da aka daidaita umarni (3–7B); auna kuma auna jinkirin UX.
• Yi amfani da Filin wasa don daidaita tambayoyi, sannan haɗa wannan tsarin a cikin lambar.
• Ƙara madadin (samfurin da ya fi girma kaɗan ko hanyar samarwa) don amintuwa.

• Gwada samfurin a cikin Filin wasa
• Samar da maɓallin API: Ƙirƙiri Maɓallin API
• Sanya ta hanyar Bayanin API

2) Ƙungiyar samfur tana ƙara taƙaitawa & hira zuwa wani app da ke akwai

• Fi son 7–14B samfura tare da tsawon mahallin; sanya akan SKU masu zaman kansu na mai samarwa.
• Ƙara lura (ƙididdigar token, jinkirin p95, ƙimar kuskure).
• Ajiye tambayoyi masu yawan amfani; kiyaye tambayoyin tsarin gajeru; yawo da tokens.

• Samfuran zaɓuɓɓuka & jinkiri: Duba Samfura
• Matakan ƙaddamarwa: Jagorar Mai Amfani

3) Masu haɓakawa da ke buƙatar fassarar kan-na'ura ko gefen na'ura

• Fara da Phi/Gemma/ƙaramin Qwen, an rage zuwa INT4.
• Ƙuntata girman mahallin; haɗa ayyuka (sake jera → ƙirƙira) don rage tokens.
• Kiyaye wani ShareAI mai bayar da sabis a matsayin mafaka don tambayoyi masu nauyi.

• Gidan yanar gizo na takardu: Takardu
• Tsarin masu samarwa: Jagorar Mai Samarwa

Girke-girke na kimantawa mai amfani (Kwafi/Manna)

Samfuran tambaya (tattaunawa vs. kammalawa)

# Tattaunawa (tsarin + mai amfani + mataimaki).

Nasihu: Ka kiyaye tambayoyin tsarin gajeru da bayyanannu. Fi son fitowar da aka tsara (JSON ko jerin maki) lokacin da za ka tantance sakamako.

Ƙaramin saitin zinariya + iyakokin karɓa

• Gina 10–50 abu saitin tambaya tare da amsoshin da ake tsammani.
• Fayyace wucewa/fallasa ƙa'idoji (regex, rufe kalmomi, ko tambayoyin alƙali).
• Bibiyar ƙimar nasara kuma jinkiri a cikin samfuran da ake nema.

Tsare-tsare & binciken tsaro (PII/alamu masu ja).

• Blocklist kalaman batanci bayyanannu da regex na PII (imel, SSNs, katunan bashi).
• Ƙara ƙin yarda manufofi a cikin tsarin tambaya don ayyuka masu haɗari.
• Tura shigarwa marasa lafiya zuwa samfurin mafi tsauri ko hanyar duba ɗan adam.

Iya lura

• Shiga tambaya, samfurin, tokens a ciki/fitowa, tsawon lokaci, mai bayarwa.
• Faɗakarwa akan jinkirin p95 da tashin hankali na token da ba a saba gani ba.
• Kiyaye wani sake kunnawa littafin rubutu don kwatanta canje-canjen samfurin a tsawon lokaci.

Tura & Inganta (Na gida, Cloud, Hybrid)

Fara da sauri na gida (CPU/GPU, bayanan ƙididdigewa)

• Ƙididdige zuwa INT4 don kwamfutocin tafi-da-gidanka; tabbatar da inganci kuma ƙara mataki idan an buƙata.
• Gudanar da fitarwa don kula da saurin UX.
• Ƙayyade tsawon mahallin; fi son sake tsara+ƙirƙira akan tambayoyi masu yawa.

Uwar garken hangen nesa na girgije (masu jituwa da OpenAI na'ura mai ba da hanya)

• Yi amfani da SDK mai jituwa da OpenAI kuma saita URL na asali zuwa wurin ƙarshen mai ba da ShareAI.
• Rarraba ƙananan buƙatu inda ba zai cutar da UX ba.
• Tafasa tafkuna da gajerun lokutan jiran aiki suna kiyaye ƙarancin jinkiri.

Daidaitawa & na'urorin haɗi (LoRA/QLoRA)

• Zaɓi na'urorin haɗi don ƙananan bayanai (<10k samfurori) da saurin juyawa.
• Mayar da hankali kan daidaiton tsari (daidaita sautinka da tsarin yankinka).
• Gwada da saitin zinariya naka kafin aikawa.

Dabarun sarrafa farashi

• Ajiye tambayoyi da mahallin da ake maimaitawa.
• Rage tsarin umarni; haɗa misalai kaɗan zuwa jagororin da aka tsince.
• Fi son ƙananan samfura idan inganci ya “isa”; tanadi manyan samfura don umarni masu wahala kawai.

Me yasa Kungiyoyi ke Amfani da ShareAI don Samfura Buɗe.

150+ samfura, mabuɗi ɗaya.

Gano da kwatanta samfura buɗe da masu masauki a wuri guda, sannan canza ba tare da sake rubuta lambar ba. Bincika Samfuran AI

Filin wasa don gwaji nan take.

Tabbatar da umarni da hanyoyin UX cikin mintuna—ba tare da kayan aiki, ba tare da saitawa ba. Bude Wurin Wasa

Takardu da SDKs guda ɗaya.

Sauƙaƙe, mai jituwa da OpenAI. Fara nan: Fara da API.

Tsarin masu samarwa (zaɓi + ikon farashi).

Zaɓi masu samarwa ta farashi, yanki, da aiki; kiyaye haɗin ku cikin kwanciyar hankali. Bayanin Masu Samarwa. · Jagorar Mai Samarwa

Ciyarwar sakin.

Bibiyar sababbin saukarwa da sabuntawa a cikin tsarin. Duba Saki.

Babu shinge Auth

Shiga ko ƙirƙiri asusu (yana gano masu amfani da ke akwai ta atomatik): Shiga / Yi rijista

FAQs — Amsoshin ShareAI Masu Kyalli

Wanne kyauta buɗaɗɗen tushe samfurin samar da rubutu yafi dacewa da amfani na?

Docs/chat don SaaS: fara da 7–14B samfurin da aka daidaita umarni; gwada nau'ikan dogon mahallin idan kuna sarrafa manyan shafuka. Edge/akan na'ura: zaɓi 2–7B ƙananan samfura; ƙididdige zuwa INT4. Yare da yawa: zaɓi dangin da aka sani don ƙarfi a cikin ba Ingilishi ba. Gwada kowanne a cikin mintuna a cikin Filin wasa, sannan kulle mai bayarwa a cikin Duba Samfura.

Zan iya gudanar da waɗannan samfuran akan kwamfutata ba tare da GPU ba?

Ee, tare da ƙididdige INT4/INT8 da ƙananan samfura. Ka kiyaye tambayoyi gajeru, aika tokens, kuma ka iyakance girman mahallin. Idan wani abu ya yi nauyi sosai, ka tura wannan buƙatar zuwa samfurin da aka shirya ta hanyar haɗin ShareAI naka.

Ta yaya zan kwatanta samfura cikin adalci?

Gina ƙaramin saitin zinariya, ayyana ka'idojin wucewa/ƙasa, kuma ka rubuta ma'aunin token/jinkiri. ShareAI Filin wasa yana baka damar daidaita tambayoyi da sauri ka musanya samfura; API yana sauƙaƙa yin A/B tsakanin masu samarwa tare da lambar iri ɗaya.

Menene mafi arha hanyar samun fassarar samarwa mai inganci?

Amfani ingantattun samfura 7–14B don 80% na zirga-zirga, ajiye tambayoyi masu yawan maimaituwa, kuma ka tanadi manya ko samfuran MoE don tambayoyi masu wahala kawai. Tare da hanyar ShareAI na masu samarwa, zaka kiyaye haɗin guda ɗaya kuma ka zaɓi mafi arha ƙarshen aiki bisa nauyin aiki.

Shin “open weights” iri ɗaya ne da “open source”?

A'a. Open weights sau da yawa suna zuwa tare da takunkumin amfani. Koyaushe ka duba lasisin samfurin kafin aikawa. ShareAI yana taimakawa ta hanyar sanya alamar samfura da haɗa zuwa bayanin lasisi a shafin samfurin don ku zaɓi da tabbaci.

Ta yaya zan daidaita ko gyara samfurin cikin sauri?

Fara da LoRA/QLoRA na'urorin haɗi akan ƙananan bayanai kuma tabbatar da inganci da saitin zinariya. Masu samar da yawa a ShareAI suna tallafawa hanyoyin aiki na tushen na'ura don ku iya yin gyare-gyare cikin sauri ba tare da sarrafa cikakken gyare-gyare ba.

Zan iya haɗa samfuran buɗe tare da waɗanda aka rufe a bayan API guda ɗaya?

Ee. Ka kiyaye lambar ka cikin tsari tare da keɓaɓɓen keɓaɓɓen OpenAI kuma ka sauya samfuran/masu samarwa a bayan fage ta amfani da ShareAI. Wannan yana ba ka damar daidaita farashi, jinkiri, da inganci a kowane ƙarshen.

Ta yaya ShareAI ke taimakawa wajen bin doka da aminci?

Yi amfani da manufofin tsarin-tsokaci, matattara na shigarwa (PII/alamu masu haɗari), kuma tura tsokaci masu haɗari zuwa samfuran da suka fi tsauri. ShareAI’s Takardu yana rufe mafi kyawun hanyoyin aiki da alamu don kiyaye bayanai, ma'auni, da madadin su kasance masu duba don nazarin bin doka. Karanta ƙarin a cikin Takardu.

Kammalawa

Samfurin mafi kyawun samfuran samar da rubutu kyauta yana ba ku saurin gyare-gyare da tushe mai ƙarfi ba tare da tsare ku cikin nauyin nauyin nauyi ba. Fara da ƙarami, auna, kuma faɗaɗa samfurin (ko mai samarwa) kawai lokacin da ma'aunin ku ya buƙace shi. Tare da RabaAI, za ku iya gwada samfuran buɗe da yawa, kwatanta jinkiri da farashi a tsakanin masu samarwa, kuma ku fitar da API guda ɗaya mai tsauri.

• Bincika Kasuwar Samfura: Duba Samfura
• Gwada tsokaci a cikin Filin wasa: Bude Wurin Wasa
• Ƙirƙiri maɓallin API ɗinka kuma gina: Ƙirƙiri Maɓallin API