Tali Models

Production model index for the Tali on-device AI runtime that powers TaliOS โ€” the voice-first, accessibility-driven operating layer built by Intelligent Devices LLC / Augmentalis Inc.

This repository is a catalogue. The actual model weights live in dedicated sibling repositories under the IDEAHQ namespace. Models are downloaded automatically by TaliOS apps via the AssetDelivery module โ€” there is no manual installation step for end users.

Model Families

Family Purpose in TaliOS Repository
Tali STT Converts spoken audio into text. Runs continuously when the user is talking to TaliOS. Tuned for low latency on mobile NPUs so commands feel instantaneous. IDEAHQ/ava-whisper
Tali NLU Embedding model for Natural Language Understanding. Maps a sentence to a fixed-size vector so TaliOS can match what the user said against the catalogue of voice commands and accessibility hints, even when phrasing varies. Runs in milliseconds on CPU. IDEAHQ/ava-nlu
Tali Storm On-device LLM (large language model) family for general dialogue, command interpretation, and tool-use. Sized for phones and tablets โ€” small enough to run without the network, large enough to hold a real conversation. IDEAHQ/ava-storm
Tali Nautilus On-device LLM family tuned for reasoning-heavy and vision-language workloads. Used when TaliOS needs to plan multi-step actions, reason over screen content, or interpret images. IDEAHQ/ava-nautilus

Catalogue (snapshot of shipped variants)

Variant Family Format Size
Tali STT Tiny STT AON-encrypted ~50 MB
Tali STT Base STT AON-encrypted ~150 MB
Tali STT Small STT AON-encrypted ~500 MB
Tali NLU 384 Base NLU embedding (English-lite) AON-encrypted ~80 MB
Tali NLU 384 Multi NLU embedding (multilingual) AON-encrypted ~85 MB
Tali NLU 768 Multi NLU embedding (multilingual, hi-dim) AON-encrypted ~140 MB
Tali Storm 2B Q4 LLM (entry-tier phone) AON-encrypted ~1.5 GB
Tali Storm 4B Q4 LLM (mid-tier phone) AON-encrypted ~2.5 GB

The full per-family variant list lives in each family's repository README.

Why on-device

Every model in this catalogue runs on the user's device. TaliOS does not send voice or screen content to a server for inference. This is a privacy-by-architecture decision, not a feature flag โ€” there is no cloud fallback path in the production runtime.

File Format

All shipped weights are wrapped in AON โ€” Tali's encrypted, signed asset container format. AON files cannot be opened without the runtime; they are decrypted in-process and cached in private app storage. The .aon extension is the only format external tooling sees.

License

Proprietary โ€” Intelligent Devices LLC / Augmentalis Inc. See augmentalis.com/license.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support