PARE achieves 52% parameter reduction on Wan2.1-14B with just 0.6 points drop on VBench, using spatial-temporal aware pruning and content-adaptive routing. Combined with step distillation, total speedup reaches 50x.
Microsoft open-sourced Lens, a 3.8-billion-parameter text-to-image diffusion model, on May 25, 2026. It rivals FLUX and SD3, runs in diffusers and ComfyUI, under MIT license.
OpenBMB's MiniCPM5-1B is a 1.08B Apache 2.0 LLM that ranks first on the Artificial Analysis index for small models, scoring 17.9 against Qwen3.5-2B's 16.3, runs on CPU, and supports a 131K-token context.
An open-source DeepSeek-native coding agent built around prefix-cache stability. A 435M-token session that costs $61 on Claude runs $12 on Reasonix.
NVIDIA Toronto AI Lab open-sources PiD, a plug-in pixel diffusion decoder that replaces VAE in FLUX, SD3, and Z-Image to output 2K-4K in one distilled pass.
Perplexity open-sourced Bumblebee, a read-only supply-chain scanner for macOS and Linux that detects compromised packages.
A solo developer just open-sourced framedex, an MIT-licensed local pipeline that indexes a year of personal video footage on a 2021 MacBook using a quantized Gemma 4 31B model running through LM Studio.
Cohere released Command A+ under Apache 2.0 on May 21, 2026. The 218B sparse MoE runs on two H100s, with native citations and 48 languages.
Tencent ARC open-sourced Pixal3D, a SIGGRAPH 2026 image-to-3D model generating high-fidelity meshes from single images.
NVIDIA released the Nemotron-Labs-Diffusion family on Hugging Face, an open-weights LLM that switches between autoregressive, diffusion, and self-speculation decoding for 2.7x to 3.3x throughput gains.
Open-source Forge framework adds guardrails to 8B local models, lifting agentic task reliability to 86.5% -- no cloud API required.
StemDeck v0.5.0 splits any track into vocals, drums, bass, guitar, piano, and other elements. Runs locally, free, no account required.
ByteDance Research released Lance, a 3B Apache 2.0 unified multimodal model that handles image and video generation, editing, and understanding in a single framework. Strong VBench and GenEval scores.
Semble gives AI agents semantic code search with 98% fewer tokens than grep+read. Indexes in 263 milliseconds. MCP-ready for Claude Code.
ComfyUI-Mesh added LTX 2.3 support on May 17, splitting the 22B video model across two GPUs over gigabit Ethernet using NVENC codec compression.
Zerostack is a pure Rust coding agent that launched May 16, 2026, running in 8MB of RAM compared to 300MB for JavaScript-based alternatives like Opencode.
NVIDIA's SANA-WM is a 2.6 billion-parameter open-source world model that generates 720p, 60-second video with 6-degree-of-freedom camera control on a single GPU.
ComfyUI-DramaBox added LoRA weight injection on May 16, letting creators load custom voice personalities into workflows without model reloads.
DeepSeek-V4-Flash is the first local model competitive with frontier AI, making LLM activation steering practical for the first time. A guide for creators.
Supertonic 3 is an open-weights, CPU-only TTS engine from Supertone with 31 languages, expression tags, and zero-shot voice cloning.
Black Forest Labs' FLUX.2-klein-9B now integrates with ComfyUI via NKD Klein Tools v1.7.0, enabling sub-second image generation with just 4 inference steps on an RTX 4090.
holaOS 0.1 ships Dashboard, Sub Agents, and Multi Workspaces for managing parallel AI workflows on macOS.
Anthropic's Claude Code ported Bun from Zig to Rust in nine days: 1,009,257 lines, 99.8 percent test pass, 13,000-plus unsafe blocks. The case study.
IBM has released Granite Embedding Multilingual R2: a pair of Apache 2.0 embedding models with 32K context, 200+ languages, and a top MTEB score under 100M parameters. A drop-in swap for paid commercial embedding APIs.