Tali Storm

Tali Storm is the general-purpose on-device LLM (large language model) family for TaliOS.

What it does

Storm is the model that holds a conversation with the user. When you say "draft an email to the team about Friday's release" or "summarize what's on this screen," Storm is what generates the answer. It runs entirely on the device — no audio, screen content, or text leaves the user's hardware to produce a Storm response.

Storm is also the engine behind TaliOS command interpretation: turning a free-form spoken sentence into a structured action that the OS can execute against an app's accessibility surface.

Where it sits in TaliOS

User speech → Tali STT → text → Tali NLU → vector match
                                     │
                                     ▼ (no exact hint match)
                               Tali Storm → action / reply

Storm only runs when faster paths (NLU vector match against the on-screen hint catalogue) cannot resolve the intent. This keeps latency low and battery cost minimal — Storm wakes up for the hard cases, not every utterance.

Variants

Each variant is sized for a specific device class. The TaliOS runtime picks the largest variant that fits the device's RAM and NPU budget at install time.

Tali ID Active Params Architecture Target
TALI-STORM-2B 2B Dense Phone (low-end)
TALI-STORM-4B 4B Dense Phone (mid-range)
TALI-STORM-26B-A4B 4B active Mixture-of-Experts Phone (flagship) / HMT-1
TALI-STORM-31B 31B Dense Desktop

Mixture-of-Experts (MoE) variants have a much larger total parameter count than active count. Only the "active" parameters are loaded for any given token, so memory cost on device is proportional to active params, not total.

Quantization: Q4_K_M for on-device targets (4-bit weights, ~75% size reduction vs full precision). Full-precision weights are retained for desktop.

File format

Shipped weights are wrapped in AON — Tali's encrypted, signed asset container. The .aon extension is the only format the OS or external tooling sees. AON files are decrypted in-process by the TaliOS runtime; they cannot be loaded by third-party inference engines.

License

Proprietary — Intelligent Devices LLC.

Downloads last month
545
GGUF
Model size
5B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support