Meta's open-weight 30B agentic multimodal model distilled from Muse Spark, designed to run always-on coding and tool-use agents locally on a single consumer.

Free open-source desktop app to run and train AI models 100% locally, with Claude Code and Codex agent integration, media generation, and private web search.

On-device work memory that captures apps, meetings, and chats every 2 seconds, then feeds that timeline into MCP-ready agents. Pro is $18.99/user/month.

Liquid AI's open-weight 2.6B dense model trained for on-device agentic workloads, with 128K context and native tool calling.

Local models plus Ollama cloud. Not a 30-minute Llama3 tutorial.

NVIDIA's efficient open-weight 30B MoE hybrid model with 3B active parameters, 1M-token context, and single-GPU deployment for local reasoning and coding.

Jeff is an open-source family of 0.8B and 2B decision models that return a calibrated probability per option in about 22 ms on a single GPU.

TensorSharp runs local GGUF models through a .NET engine, CLI, browser chat, and Ollama or OpenAI-compatible APIs.

Tencent’s model compression toolkit combining quantization, distillation, sparse attention and speculative decoding workflows.

An efficient local LLM inference library built around the EXL3 quantization format, running large open models on consumer GPUs with tensor and expert parallelism, and powering the TabbyAPI server.

halogen-flash-server is a Strix Halo inference server that runs Qwen3.8-Flash-Next with an OpenAI-compatible API.

mlx-serve is a MIT Zig inference server for Apple Silicon that runs MLX and GGUF locally with OpenAI and Anthropic APIs, no Python.

NInfer is an Apache-2.0 C++/CUDA engine for selected Qwen checkpoints on a single RTX 5090, with OpenAI- and Anthropic-compatible local APIs.

oMLX is an Apache-2.0 Apple Silicon inference server with SSD KV cache and a macOS menu bar app for coding agents.

Open-source local inference engine for Qwen3.8-Flash-Next on Windows and Linux, with a chat interface and compatible agent APIs.

AMD ROCm's NPU-first runtime for Ryzen AI laptops: run LLMs, VLMs, and audio models locally with no GPU, 256K context, and a 17MB lightweight CLI.

An open-source desktop workbench for the local Grok Build CLI, with multi-project spaces, real-time agent streaming, and an integrated file and code loop across macOS, Windows, and Linux.

BaoCut combines local transcription, bilingual subtitles, translated dubbing, and agent-assisted video editing on Apple Silicon Macs.

Inco AI's Apache 2.0 block-diffusion drafter for speculative decoding. Released 2026-08-18 for Qwen3.8-27B and Muse Glimmer.

K2-Horizon-7B is an open-weight IFM language model with a 512K context window, tool calling, and released training checkpoints.

Poolside's open-weight 118B MoE coding model with 8B active parameters, a 1M-token context window, native interleaved reasoning, and an OpenMDW-1.1 license.

Ling-3.0-flash-VL is inclusionAI's MIT multimodal MoE: 124B total, 5.5B active, with image and video.

inclusionAI MIT hybrid-linear MoE: 124B total, 5.1B active.

Ling-3.0-tiny is inclusionAI's MIT hybrid-linear MoE: 7.9B total, 1.3B active, for local agents.