IBM released Granite 4.1 on April 29, 2026 with dense decoder-only 3B, 8B, and 30B Apache 2.0 LLMs that walk back the Granite 4.0 MoE bet. The 8B dense beats the 32B-A9B MoE on most benchmarks. 128K production context, 512K extension.
Zed Industries shipped Zed 1.0 on April 29 2026, a Rust-built, GPU-accelerated code editor with parallel agents, edit prediction, and the open Agent Client Protocol.
Mistral shipped Medium 3.5 as a 128B dense multimodal model under modified MIT, with cloud Vibe agents and Le Chat Work mode on day one.
NVIDIA released Nemotron 3 Nano Omni on April 28: a 30B open-weight model that handles text, image, video, and audio in one architecture, with joint audio-visual reasoning and multi-hour context.
ComfyOrg announced ComfyStudio on April 28: a Los Angeles VFX and animation residency that pairs working artists with the ComfyUI team and releases every workflow open-source.
Poolside released the first two models in its Laguna family on April 28: a 225B-parameter flagship called M.1 and an open-weight 33B model called XS.2 under Apache 2.0. Both are pitched at agentic, long-horizon coding work.
SenseTime open-sourced SenseNova U1 on April 28, 2026, releasing three model variants under Apache 2.0. The architecture drops VAEs entirely, unifying image generation and text reasoning in one space.
The complete creator guide to ComfyUI in 2026: setup, models, partner nodes, working workflows for image, video, audio, and 3D, the ecosystem, and managed-vs-self-host trade-offs.
The complete creator reference to open-source AI in 2026: LLMs (DeepSeek V4, Kimi, Qwen), image (FLUX, Wan 2.7), video (Wan, Skywork, LTX), 3D (HY-World, TRELLIS), audio (Voicebox, OmniVoice, Darwin-TTS), all license-aware.
ComfyUI v0.20.1 lands April 27 with SUPIR upscaling, RIFE and FILM frame interpolation, SAM 3.1 segmentation, plus 4K Veo and Kling and GPT-Image-2 partner nodes.
DeepSeek V4 Preview ships MIT-licensed 1.6T and 284B MoE models with 1M context at sub-$0.30 per million output tokens. What it means for creators.
Eyeline Labs released Vista4D, an open-source framework that reshoots existing video from new camera angles using a 4D point cloud plus video diffusion.
Inclusion AI released LLaDA2.0-Uni, a 16B MoE diffusion model that unifies text-to-image generation, instruction-based editing, and image understanding in one Apache-2.0 checkpoint.
Alibaba released Qwen3.6-27B on April 22, 2026. The 27B dense open-weight model beats its larger 35B-A3B sibling on coding, agent, and vision benchmarks.
Xiaomi released MiMo-V2.5 on April 22, an 8B open-weights speech recognizer plus a three-model TTS series with prompt-built voice generation.
Alibaba's Taobao and Tmall algorithm team released Tstars-Tryon 1.0 on April 21, 2026, a commercial-scale virtual try-on system already serving millions of Taobao shoppers.
Jamie Pine's Voicebox brings seven TTS engines and voice cloning to your local machine. Free, open source, and entirely offline.
Alibaba released Qwen3.6-Max-Preview on April 20, 2026, topping six programming benchmarks including SWE-benchPro. Here is what it means for creators building AI workflows.
ComfyUI added Quiver Arrow 1.1 as a Partner Node on April 20, 2026, bringing structured SVG generation to open-source visual workflows for the first time.
Moonshot AI's Kimi K2.6 scores 58.6 on SWE-Bench Pro, placing first above GPT-5.4 and Claude Opus 4.6. Full architecture breakdown, benchmark analysis, and what the open-weight license actually allows.
A community port of Microsoft TRELLIS.2 brings image-to-3D mesh generation to Apple Silicon. No Nvidia GPU required, GLB output ready for Blender and Unity.
Researchers published Darwin-TTS on April 15, 2026, a text-to-speech model that adds emotional expression to AI voice without any training, fine-tuning, or new data.
HeyGen launched a developer platform with an open-source CLI, an HTML-to-video rendering framework called Hyperframes, and agent skills for Claude Code, Codex, and Cursor.
NucleusAI released Nucleus-Image on April 14, a 17-billion parameter diffusion model that activates only 2 billion parameters per image -- cutting compute without sacrificing quality.