Apache 2.0 licensed models, published research, and open-source projects from Eulogik. Built for edge deployment, regulated industries, and quantitative systems.
Byte-level text classifier that runs at <6ms with a 200KB ONNX export. Published on ArXiv (2608.14658). Designed for edge deployment, air-gapped systems, and high-throughput classification pipelines.
Zero-shot time series foundation model with DeltaNet + LongConv architecture. Probabilistic forecasting with quantile regression and uncertainty quantification. Runs on Raspberry Pi.
A 256M-parameter vision-language model for document understanding. Handles OCR, layout analysis, and visual QA — all without a GPU. Built for regulated industries that process high volumes of documents privately.
A federated LLM router with human-interpretable routing logic. Reduces inference costs by 40–70% by intelligently routing requests across model tiers. Privacy-preserving by design.
A token-minimal, platform-agnostic programming language designed for AI agents. 4× more compact than Python for agent logic. Available on PyPI, npm, and featured on Product Hunt.
Edge-first Hindi/Hinglish LLM powered by our proprietary Brahmi Tokenizer. 33–45% fewer tokens on Indic text with zero byte fragmentation. GGUF, WebGPU, llama.cpp, and tool-use reasoning. 100% offline and air-gapped.
Brahmi Neural Codecs: ultra-low bitrate (<12kbps) acoustic neural speech codec engineered for Indian phonetic inventories, retroflex consonants, and constrained IoT edge bandwidth.
Indic multilingual speech recognition with frozen Whisper encoder and language-specific LoRA experts. Supports Hindi, Tamil, Telugu, Bengali, and Marathi with 20× hallucination reduction. ONNX quantized for edge deployment.
Cognitive Resonance Network adapter on frozen Gemma-4. Memory-augmented generation with EpisodicMemory retrieval table achieving 100% exact factual recall. Zero hallucination on verifiable facts. Runs on Apple Silicon via MPS.
Second-generation Cognitive Resonance Network with error-correction and logit adjustment on frozen backbone. Parameter-efficient capability preservation. Optimized for on-device AI on Apple Silicon.
Calibrated banking decision model. One forward pass on a state + typed question returns calibrated probabilities. 87.9% on Banking77 test (3,080 cases), beating Jev API zero-shot (80.3%) by +7.6 points. Apache 2.0. Runs fully local, no API key, no per-call cost.
Sub-1B agentic function-calling model on MiniCPM5-1B with decoupled task routing (DTSA). Achieves 3.52× baseline on ToolACE agentic benchmarks. 4-bit GGUF for on-device deployment. 690+ downloads.
149M ModernBERT cross-encoder reranker for high-precision RAG. Reranks retrieval results in <50ms on CPU with an 8K context window. Eliminates hallucinations in legal and financial document search.
Standard Western tokenizers (GPT-4o, LLaMA-3) lack native Devanagari subwords, fragmenting Indic scripts into raw UTF-8 byte sequences. Indian enterprises pay up to 4× higher API costs and lose over 50% of effective context length. Eulogik’s Brahmi Tokenizer preserves morphological aksharas with surgical compression.
Fragmented into isolated characters and multi-byte UTF-8 codepoints. High latency and 3-4× billing inflation.
Surgical Devanagari subwords preserved. 0 byte-fragmentation, +34% effective context, and ~2.2× faster inference.