Sarvam Prism: India ka naya open-weights 'decision model' — 44 ms me jawab, Hindi ke saath

Bengaluru ki AI company Sarvam ne 3 November 2026 ko ek naya model release kiya hai: Sarvam Prism. Ye ChatGPT jaisa chatbot nahi hai. Ye ek "decision model" hai, yaani aisa model jo kam time me tez aur sahi faisle (jaise kisi ticket ko kis category me daalna, kaun sa option chunna, haan ya na) lene ke liye bana hai. Neeche saari baatein Sarvam ke apne launch blog se hain, aur benchmark numbers Sarvam ne khud test karke diye hain, kisi independent lab ne nahi.

Prism kya hai?

Sarvam ke mutabik Prism unke Sarvam 105B base model par bana "System One" decision model hai. System One naam Daniel Kahneman ki soch se aaya hai: tez, intuitive soch jo hamare roz ke zyadatar chhote faisle sambhalti hai. Company me bahut saare kaam aise hi chhote faisalon ki chain hote hain, aur wahan har baar bada chatbot chalana mehenga aur dheema padta hai.

Sarvam ne kya claim kiya

• Sarvam ne 25 decision models ko apne tareeke se test kiya. Prism ka average score 71.8 raha aur uska median response time 44 millisecond (ek NVIDIA B200 GPU par).
• Blog ke mutabik 100 ms se kam time me jawab dene wale models me Prism ka score sabse zyada hai. Kuch models ka score isse zyada hai, lekin wo 2 se 6 guna zyada time lete hain.
• Indian languages ke test me (10 Indian bhasha aur Hinglish) Prism ka score 82.5 hai. Sarvam ke hisaab se ye test me shamil sabse achhe models ke barabar ya behtar hai.
• Calibration: jab model 90% ya zyada confident hota hai, tab English me wo 3.3% galti ka andaza lagata hai aur asal me 2.2% galti karta hai. Indian languages me andaza 5.4% aur asal galti 4.4% hai. Matlab model apni galti ko kam nahi aankta, aur company use confidence ke hisaab se kaam automate kar sakti hai.

Ek zaroori baat: ye saare numbers Sarvam ke apne tests hain. Production me use karne se pehle apne data par khud check karna chahiye.

Open weights, Apache 2.0

Sabse badi baat ye hai ki Prism open weights ke saath aaya hai. Model Hugging Face par sarvam/prism naam se Apache-2.0 license ke saath dikh raha hai, yaani startups apne data par isko train karke apne server par chala sakte hain. Hugging Face page par Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Odia aur Punjabi languages tag hain.

Hindi bolne wale students aur freelancers ke liye matlab

• Agar tum AI seekh rahe ho to ab "sirf chatbot" se aage dekho: classification, routing, approval jaise kaam me chhote tez models ka use badh raha hai. Ye skill resume me kaam aati hai.
• Indian-language support ka matlab hai ki Hindi/Hinglish customer messages ko samajhne wale tools banana ab aasaan hota ja raha hai. Freelancers aur small agencies ke liye ye naya kaam ka mauka hai.
• Open weights ka fayda ye hai ki bina kisi vendor ke lock-in ke tum experiment kar sakte ho. Shuru karne ke liye Hugging Face model card padho aur chhote dataset par try karo.

Pichle hafte ki ek aur Sarvam khabar

27 October ko Sarvam ne HQ IDS (Ministry of Defence ke tahat ek sanstha) ke saath ek MoU ka elaan kiya, jisme Army, Navy aur Air Force ke liye AI projects ka umbrella framework banega. Sarvam ne 28 October ko saaf kiya ki MoU HQ IDS ke saath hua hai, seedhe Ministry of Defence ke saath nahi. Details Sarvam ki announcement me padh sakte ho.

Aivesh Academy par next kya?

Hum Aivesh Academy par AI ko Hindi me seekhne ke free resources banate rahenge. Agar tum chahte ho ki hum Prism jaise kisi model ka step-by-step "kaise use karein" guide banayein, to WhatsApp channel par bata do.

Source: Sarvam Prism launch blog, Hugging Face model page.

— Devesh Sarkaar, Founder, Aivesh Academy

Join our WhatsApp channel: https://whatsapp.com/channel/0029Vb8x2347YSd7f4EUue3q