← Chitta

Model map

Chitta 0.8.7 names every specialist in Sanskrit. Familiar icons show its role. Models and Downloads keep technical model names hidden unless you intentionally reveal the details.

Models download separately with your choice. The app shows download size, resumable progress and file verification. Optional models are not bundled with the installer. Final model quality and performance checks for this release remain pending.

Hover a markMeanings appear here. In the app: Settings → Model map.

More choices in 0.6

Śilpī-Medhā maps GPT-OSS-20B; Śilpī-Dakṣa and Śilpī-Dakṣa-Laghu map Qwen3.6 quantizations. The Gemma-based Śilpī-Mūla family remains available. Buddhi begins with task requirements and published evidence, then Prāṇa checks resources. Feedback refines exact model rankings; this is not a claim of newly trained model weights.

Citra-Akṣara is an optional lettering-focused image stack with a 14.97 GB complete download. Actual runtime compatibility, quality and 16 GB VRAM fit remain unverified. Review model licensing and download size in Chitta.

Versions, previews and limitations

Your device, and the memory available now

Total system RAM and verified dedicated GPU memory guide the models Chitta offers. Before a model loads, Prāṇa rechecks the RAM and VRAM available at that moment, along with the active runtime. Shared GPU memory is not counted as extra RAM.

Prāṇa can release eligible idle Chitta models. Closing another application requires your review and confirmation that your work is saved. These are planning estimates and protective checks, not a guarantee that a model will fit or run quickly on your hardware. The release records bounded owner-machine tests; performance and fit on other hardware remain unverified.

Buddhi, Prāṇa and Karma in 0.8.0

Buddhi keeps existing neural routing; the optional Mārga (Arch) router now has its own small slot and no longer stalls a turn. Prāṇa uses measured per-model memory, speed and load time from Device profile calibration; a blocked work turn steps down to Vāk with a note. Karma retries an empty coder answer once, then hands it to a smaller model.

0.8.0 changes and measured limits

Also on aavahanam: Eagle Communicator keeps professional chat with optional on-device translation — complementary to Chitta, not replaced by it.

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