On August 19, 2026, Ornith AI released Ornith-1.5 as open weights under the MIT license: a 397B mixture-of-experts flagship, a 35B MoE that activates 3B parameters per token, and a 9B dense model with a quantized mobile build.
On August 19, 2026, TrueFoundry released TrueForge, an open source agent harness under the MIT license on GitHub and PyPI. The company claims 30 to 75 percent lower cost per completed task, achieved through context engineering rather than a cheaper model.
On August 18, 2026, Comfy Org open sourced Comfy MCP on Local, letting an AI agent connect to and drive a ComfyUI installation running on your own machine. The agent reads your GPU specs, sees which models and custom nodes you have, and builds workflows from a plain-language description.
Z.ai shipped GLM-5.3 on August 14, 2026, reusing the GLM-5.2 base and pushing large coding gains through scaled post-training. It is live via API now, with open weights in about two weeks.
MiniMax released Music 3 as open weights under Apache 2.0, and Comfy-Org repackaged it for ComfyUI the same day. It generates full five-minute songs from lyrics plus a description.
DeepSeek Harness is an open, MIT-licensed framework for building AI agents where every component is a plugin. Here is what shipped and how to run it.
DeepSeek pushed its flagship V4 Pro to general availability on August 12, 2026, ending a preview that ran nearly four months. The GA build is a 1.6 trillion parameter MoE with a 1M-token context window.
A new Apple Silicon fork of Automatic1111 drops local Stable Diffusion render times from 8 to 10 seconds to 3 to 7 seconds on an M3 Pro, using native Metal kernels while keeping the standard interface.
Xiaomi's MiDashengLM-Gen is an open-weights 3B model that generates speech, music, and sound effects together from a single text prompt. Here is what it does and how to run it.
MiniMax H3 gave open-weights creators a 15-second 2K omni-modal video model. H3 Studio collapses the whole ComfyUI pipeline into a single node and a single page, with a local prompt LLM and automatic character consistency.
Unsloth AI released Unsloth Desktop, a free open-source app that runs and fine-tunes LLMs and diffusion models locally, 2x faster with 70 percent less VRAM.
Removing a background from video used to mean a green screen and hours of rotoscoping. In 2026 you can pull a clean cutout from ordinary footage in minutes, no green screen, using AI segmentation that runs locally or in a browser.
NVIDIA released Nemotron 3.5 Lightning, a 30B open-weights MoE model for always-on agents, alongside NeMo Switchyard, an open-source router that cuts agent cost by escalating only the hard steps to a frontier model.
Say It, a free and open-source macOS app, turns highlighted text into spoken audio and clones voices entirely on your own machine, with no cloud or accounts.
NVIDIA released Magpie TTS Multilingual, an open-weights text-to-speech model that speaks 12 languages and hits a 32ms time-to-first-audio, small enough to self-host inside a live voice agent.
Meta released Muse Glimmer, a 30-billion-parameter open-weights agentic model built to run on a single consumer GPU. Announced August 10, 2026 under Apache 2.0, it is distilled from Muse Spark and ships on Hugging Face.
Vibez is a new open-source DAW written in pure Rust, launched August 9, 2026 for Linux, macOS, and Windows, pairing clip launching with a timeline plus VST3 and CLAP plugin support.
PrunaVAED, a new open-weight decoder from Pruna AI, is a drop-in replacement for the LTX-2.3 VAE that decodes video up to 2.08x faster and cuts peak VRAM roughly in half.
Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter open-weights agentic model that runs entirely on-device, planning and calling tools with no cloud API cost.
CopilotKit open-sourced the Channels SDK, a MIT-licensed toolkit for dropping AI agents into Slack, Teams, Discord, and Telegram with native interactive UI.
Cloudflare has released Cloudflare OS, an open-source platform for building and running AI agents and full-stack apps, available today on GitHub and deployable into your own Cloudflare account in minutes.
Which local fine-tuning tool should you install? Soup, Unsloth, Axolotl, and LLaMA-Factory compared on the constraints that actually decide it.
Alibaba Tongyi Lab released Wan-Animate-2 on August 7, a 14B Apache-2.0 character animation model. Two days later it landed in ComfyUI as native nodes.
Prime Intellect released Prime Agent on August 5, 2026, an open-source coding harness that rewrites its own instructions while it works. Paired with Opus 5, it beat the human-expert baseline on ARC-AGI-3.