Model card: Jivra 1

Version 1.7b-q4km-2026.10.24 · Chundu Laboratories · Updated 24 October 2026

Overview

Jivra is a general-purpose AI assistant for everyone: everyday questions, explanations, writing help, maths, code and Indian topics, in English, Hindi and Telugu.

Model details

• Base: Qwen3-1.7B open-weight model (Apache-2.0 licence), quantised to 4-bit (GGUF Q4_K_M) to run on CPUs.
• Served with the open-source llama.cpp server on Chundu Laboratories' own machine. No closed or third-party AI services are used.
• Context window: 4,096 tokens. Replies are capped at 512 tokens.
• A same-size backup copy answers when the main copy is down or busy.

How we test it

Before any new version replaces the current one, it must score at least as well on two held-out test sets it was never trained on: 44 maths questions with exact answers, and 40 general questions (advice, writing, coding, general knowledge, Indian topics, safety, Hindi and Telugu). On the general set, the live version passed 35 of 40 (87.5%) in our October 2026 test.

Intended use

• Learning, everyday help, drafting and brainstorming.
• Not for medical, legal or financial decisions, emergencies, or anything where a mistake could cause harm.

Limitations

• It is a small model. It can be wrong with confidence, especially about facts, dates and names. Check important answers.
• Knowledge is limited to its training data; it cannot browse the web.
• Hindi and Telugu answers are weaker than English.
• Replies can take from a few seconds to a few minutes because it runs on CPUs.

Safety

• One shared safety filter covers chat, images and videos and blocks sexual content involving minors, violent hate and dangerous instructions. Images and videos also block nudity, gore and pictures of real people.
• Jivra's instructions tell it to refuse harmful requests and to respond with care to people in distress.
• Keys that repeatedly break the rules are paused.

Data and privacy

Prompts and replies are not stored. Request logs hold only timing, status, token counts and a short key prefix.

Questions: see Contact.