Daily AI.
No monthly bill.

Windows 0.9.11: Arch-Router's waiting time has one owner set from how long routing really takes on this PC, and a timed-out routing request no longer blocks later turns; pictures are planned against what really finished on this PC and a memory stop is reported as low memory; a stored coding-model memory peak far above what the model file can need is set aside for recalibration, with the 2 GB floor unchanged. Includes 0.9.7, 0.9.8, 0.9.9 and 0.9.10, which were not published on their own: Arch-Router routes again, programs are checked by running them, multi-part programs are planned instead of declined, and a chat keeps its context when it moves between models. Everything in 0.9.6 stays: Chitta can now use a plugin you install yourself in C:\AI\Chitta\plugins (Chitta never ships one): after you agree once, matching questions are answered by it in the chat, Prāṇa checks memory before it runs, and keys stay in Windows secure storage; plugin examples are prepared in the background so they never hold up chat, pictures or web answers; and Arch-Router's trained adapter now really loads in the app. Includes the 0.9.5 plugin slot, which was not published on its own. Everything in 0.9.4: Arch-Router answers with a short route name using a small Chitta-trained adapter, so routing is quicker (a quick check on one Surface PC: about 0.25 s instead of 0.74 s for the routing answer); a neural chain check keeps only clear single requests away from Arch-Router; Rakṣā locks face pixels only when you ask and finds the faces itself; photo-reference pictures use the strongest model that can take a reference and say why; first-run setup selects Buddhi · Mārga, Cakṣu, Smṛti and Nyāya by default; and the 🎮 games orb is always at the bottom-right. Everything in 0.9.3: Requests are routed by meaning: a quick neural check for clear asks and Arch-Router 1.5B for longer, pasted or multi-step ones, with requests that need real-world facts starting with web research; Vāk never routes or takes picture, code, research or PC work; and each answer has Regenerate and Re-route buttons. Includes the 0.9.2 speed-aware model choice, which was not published separately. Everything in 0.9.1: Writing, explanations and careful questions now use the strongest model that fits instead of the quick 2.6B model; nothing is loaded when Chitta starts; a pasted prompt is read as your request; naming a model for one step applies to that step; and Prāṇa has a clearer memory view that trims idle background apps (nothing closed) before choosing a smaller model. Everything in 0.9.0: Coding work stays on Śilpī with its full working context and never drops to the small 2.6B quick model while a better model fits; naming a model ("handover to Citra Uttama") always uses it, with the last file handed over; Prāṇa shows a live memory panel and can trim idle background apps (opt-in); documents are created again, streamed text stays visible, and computer control always asks first and hands back. Everything in 0.8.9: The retry card now appears when the check finds something to improve, with your answer kept; an honest "couldn't finish" no longer shows a passing review. Everything in 0.8.8: Simple asks answer fast; each answer is checked once and you choose whether to retry with the suggested improvements (Deep and Thinking hand you an improved prompt); coding keeps only the best version of a file; an honest closing message on low memory that names any files written; a smoother progress panel; optional Hasta computer helper (UI-TARS) downloadable and verified from Downloads; click the Chitta name or logo for About Chitta (version, credits, thanks to Sadhguru). Changes and limitations.

Application update for Chitta 0.9.6, 0.9.4, 0.9.3, 0.9.1, 0.9.0, 0.8.9, 0.8.8, 0.8.7 and 0.8.6 - 177.8 MB. Apply through Chitta's update check; do not run this file. Other installations need the full installer.

Models download separately after your review. Windows is unsigned; Android is 0.8.8. Release notes - Checksums

Made by R Prabhakar · rs.prabhakar@gmail.com