Sorika Foundation Models
Sovereign, ultra-efficient speech, multimodal, and reasoning architectures engineered by Darsh Yadav. Built to deliver State-Of-The-Art accuracy with architectural frugality on consumer and community hardware.
Active Foundation Engines
Vaani-2 Flash ASR
High-Throughput Speech Recognition (809M Turbo)#1 Overall English Accuracy (5.90% Avg WER across 7 benchmarks) with an incredible 21.5x to 30.3x Real-Time inference speed. 100% open-source permanently fused safetensors on Hugging Face.
Vaani-1.1 Lite ASR
Multimodal Speech-LLM (~600M Conformer + Qwen3)#1 SOTA on Indian Speech (11.78% Hindi Avg WER). Powered by an LLM reasoning decoder, excelling in noisy acoustic environments (3.60% on LibriSpeech Other). 100% open-source on Hugging Face.
Vaani-1 Lite ASR
Edge & Mobile Friendly ASR (244M)Slashes Hindi WER from 47.77% down to 19.32% (over 2.5x error reduction) while retaining full 99+ global languages. Runs natively on CPUs and mobile devices. 100% open-source on Hugging Face.
Suzune S1 Voice
Non-Autoregressive Neural TTS (80M)Ultra-lightweight 80M non-autoregressive voice foundation model with sub-20ms latency and 24kHz studio HD audio. Constant ~320MB VRAM footprint with zero phonetic hallucinations.
Swen 1
Adaptive Metacognitive Architecture (Low · Med · High Modes)Sorika's flagship conversational reasoning foundation architecture. Features dynamic test-time compute with configurable Low (ultra-fast), Medium (balanced multimodal), and High (deep architectural thinking) modes.
Upcoming Architectures in Training
Next-generation models currently being researched, trained, and benchmarked by Darsh Yadav at Sorika Labs.
Vaani-2.1 Flash
Targeting sub-10% Hindi WER and native Hinglish code-switching without compromising the signature 25x+ RTF inference velocity.
Swen Series
Next-generation compact spatial controller & edge reasoning foundation models engineered with hybrid conv-attention backbones and constant O(1) memory complexity.
Next-Gen Vision Model
End-to-end multimodal perception foundation capable of zero-shot document hierarchy synthesis, historical manuscript parsing, and continuous visual stream understanding.
Nano Model
Ultra-compact sub-100M parameter edge foundation model optimized to run entirely inside CPU L3 cache or embedded micro-silicon with microsecond latencies.
Thinking Fast Models
Deliberative reasoning engines with autonomous test-time compute scratchpads and constrained-entropy sampling for instant multi-step deduction and symbolic proofs.
Haruka Discovery
High-throughput neural discovery engine for multimodal real-time horizon retrieval across dense vector spaces.
Sorika Logic Proofs
Deterministic multi-step reasoning foundation engine engineered for verifiable symbolic proofs and mathematical verification.