Tag

Runs huge mixture-of-experts LLMs like DeepSeek-R1/V3 on a single 24GB GPU plus CPU DRAM by keeping attention on the GPU and offloading expert weights to CPU. Reports 3-28x speedups via Intel AMX/AVX512 kernels and fits 139K context in 24GB VRAM.

SurfSense

Connects any LLM to internal knowledge sources and lets teams chat with cited, RAG-style answers. Notable for broad connectors (Drive, Notion, GitHub, YouTube), universal LLM/embedding support, and self-hostable Docker deployment — aimed at teams that need private, searchable LLM-backed knowledge.

OmniParser

Turns a UI screenshot into structured elements so a vision LLM can act without HTML or accessibility trees. A fine-tuned detector finds interactable icons; a caption model describes their function, lifting GPT-4V grounding on ScreenSpot and Mind2Web.

bitnet.cpp (BitNet)

Official inference framework for 1-bit and ternary (1.58-bit) LLMs such as BitNet b1.58, with optimized CPU kernels. Delivers 1.37x-6.17x speedups and 55-82% lower energy on x86 and ARM, and runs a 100B model on a single CPU at 5-7 tokens/sec.

Speech To Speech

Chains four swappable open modules — voice activity detection, speech-to-text, an LLM, and text-to-speech — into a local voice agent that needs no proprietary APIs. Runs on CUDA, Apple Silicon, or Docker, with an OpenAI-compatible realtime WebSocket mode.

Karukan

Provides a neural kana→kanji Japanese input method for Linux and macOS using GPT-2–style models run via llama.cpp; offers live conversion, context-aware suggestions, learning cache, and Sudachi-based system dictionary support.

Developer framework for building AI agents that autonomously trade on Polymarket prediction markets. Bundles the Polymarket and Gamma APIs, a Chroma RAG layer that pulls in news, and a CLI to query markets, reason with an LLM, and execute trades.

NexaSDK

Runs local LLM, vision-language, ASR, OCR, and image-generation models across NPU, GPU, and CPU from one command. Differs from Ollama and llama.cpp with first-class Qualcomm Hexagon NPU support and day-0 coverage of new models like Qwen3-VL.

Trains a sub-100M-parameter LLM from scratch — pretraining, SFT, LoRA, DPO/RLHF, and distillation, sized from ~26M up to ~100M-plus dense and MoE. Headline figure: the ~64M minimind-3 variant's SFT stage runs 1 epoch in ~2h and ~3 RMB on one NVIDIA 3090.

ReMe

Stores agent memory as human- and agent-readable Markdown files with wikilinks instead of an opaque vector DB. Auto Memory/Resource/Dream jobs distill conversations into long-term notes, and hybrid wikilink + BM25 + embedding search retrieves them.

Fincept Terminal

Desktop finance analytics terminal that combines CFA-level models, real-time trading and 100+ data connectors with embedded Python for analytics; includes 37 AI agents and local/multi-provider LLM support for automated research and decision workflows.

MiniMind-V

Trains a 65M-parameter vision-language model from scratch in ~2 hours on one RTX 3090, about 3 RMB (~$0.40) of GPU rental. Connects a frozen SigLIP2 encoder to a small MiniMind LLM via a two-layer MLP projector; full PyTorch code for pretraining and SFT.