Intelligence, thoughtfully built.
An independent multimodal research studio. We architect ultra-compact, sovereign foundation models that perceive, reason, and speak with extreme mathematical frugality.
The Season of Sorika.
Nuanced. Atmospheric. Sovereign.
Acoustic intelligence.
Sub-20ms streaming latency.
Want to use this API?
Try in StudioArchitectural discipline. Built for sovereign capability.
Vaani. 100% open-source speech intelligence on community hardware.
Sorika Labs has open-sourced the entire Vaani Speech AI series on Hugging Face (huggingface.co/sorika-labs). Trained and converged entirely on community hardware — Google Colab and Kaggle community GPUs — Vaani proves that State-Of-The-Art foundation intelligence can be built without multi-million dollar cloud clusters.
Through High-Rank LoRA ($r=64/128, \alpha=128/256$), gradient accumulation, custom loss masking, and frozen encoder shielding, Vaani achieves #1 SOTA on Indian speech (11.78% Hindi WER) in the Lite Series, and #1 English accuracy (5.90% WER) with blistering 21.5x–30.3x RTF in the Flash Series. All weights are 100% permanently fused safetensors — zero adapter overhead, instant plug-and-play.
Reverie. Designed for deep conversation & creation.
Reverie is Sorika’s flagship conversational intelligence workspace — combining deep step-by-step reasoning, real-time web search, precision tool calling, and interactive diagramming.
Engineered like DeepSeek for nuanced dialogue and structured problem solving, Reverie excels at live information grounding, dynamic diagrams, and rapid prototyping of lightweight websites and interactive web games without unnecessary fluff.
Architectural frugality over brute-force compute.
The global AI landscape has centered on multi-billion parameter models and massive compute clusters. At Sorika, our thesis is fundamentally opposite: intelligence is an architectural property, not simply a measure of electrical power.
We engineer compact sub-100M parameter foundation models with constant O(1) bounded memory footprints. By innovating on novel non-autoregressive topologies, neural vocoders, and continuous representations, we achieve frontier performance on consumer silicon and local hardware without multi-million dollar cloud budgets.
Sovereign intelligence, built from India.
True technological independence requires owning the foundational cognitive layer. We train our speech, audio, and reasoning models directly on the acoustic phonetics, linguistic nuances, and cultural reality of 22 scheduled Indian languages.
By maintaining sovereign architecture from day zero, Sorika ensures complete data custody, zero external inference dependencies, and deployable weights for enterprise and public infrastructure worldwide.
Latest research
View all
Suzune S1: An 80M Parameter Non-Autoregressive Speech Foundation Engine
Latest news
View allSorika Labs Releases Vaani Speech AI Series: Zero-to-SOTA on Community GPUs
Sorika Labs Accelerates Active R&D on Swen Series, Vision, Nano & Fast Thinking Models