AI study assistant for CUET, Banking, and Commerce students, powered by the Tiny AI model
Overview of self-supervised learning of tiny models, including distillation-based methods (aks. self-supervised distillation) and non-distillation methods.
⚡ A talking desktop Pikachu that runs 100% on-device — MiniCPM5-1B brain + VoxCPM voice + Nemotron ears, every model ≤1B params. No cloud, works with the Wi-Fi off.
🔍 Explore recursive reasoning with TinyRecursiveModels, a compact 7M parameter neural network achieving high scores on tough tasks without massive resources.
A deterministic proving ground for tiny local models that repair Linux hosts through a single bash tool: 172 sandboxed scenarios, observable scoring, no LLM judge.
First 1-bit (BitNet b1.58) recursive reasoner for Sudoku-Extreme - distilled from a 7M-param FP TRM teacher into a 1.4 MB ternary student
🚀 Streamline parallel development with Ralph MCP: run multiple PRDs in isolated workspaces, ensuring quality and efficient merging.
Typed-decision engine: choice, score and yes/no decisions over text in one forward pass. 16.9M params, champion test macro 0.5339 (measured). Install from git, fetch_weights, myna-serve.
Test-Time Training deployed on an STM32 NPU — self-adapting ResNet from PyTorch to Cortex-M55 edge hardware.
A curated list of tiny open-source models you can actually run — verified checksums and measured latency via tinymodels.co
Testing the Moondream tiny vision model
AI document generator powered by the Tiny AI model. Creates Word, PDF, PPTX, XLSX, and Markdown files completely client-side with real-time response streaming.
Local, project-scoped memory system for LLMs with evidence-based truth validation. Provides reliable long-term context via OpenAI-compatible Proxy and MCP, using Chain-of-Verification (CoVe) to eliminate hallucinations and the Ralph Loop for autonomous codebase repair.
A Claude Code harness for building tiny task-specific models. Every experiment must beat a baseline to ship.
A playground for pushing local AI agents as far as possible on edge/mobile and local devices.
Phi-3-Vision model test - running locally
Paired 15M experiment: full-attention vs 3:1 GatedDeltaNet hybrid, trained from scratch on Windows — hybrid wins quality, 4x smaller decode cache, no speed win at this scale
Flip-2 by Keethesh Labs. Our most capable decision model. 2 parameters. Now supports maybe.
Train the smallest LM you can that fits in 16MB. Best model wins!
🚀 Implement the Tiny Recursive Model (TRM) for improved performance in recursive tasks, building on the HRM framework by Sapient AI.
To associate your repository with the tiny-models topic, visit your repo's landing page and select "manage topics."