New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
ByteDance just open-sourced Helios, a 14B video generation model that hits 19.5 FPS on a single Nvidia H100 GPU. It can churn out minute-long videos supporting text, image, and video inputs without quality drifting. Why it matters: bringing near-real-time speeds to minute-scale generation gets open-source video much closer to interactive production.
Paperclip just open-sourced an orchestration framework that organizes AI agents into a full corporate org chart. You act as the board chair setting goals and budgets, while an AI CEO delegates tasks down the ladder using heartbeat check-ins and ticketing. While running a fully autonomous business is still closer to a simulation game, its governance controls offer a practical model for solopreneurs.
Showlab just open-sourced Kiwi-Edit, a video editing framework that pairs a multimodal LLM with a 5B diffusion transformer to edit clips using plain text or reference images. It handles 720p style transfers, background swaps, and adding or removing objects while keeping temporal consistency. With code, weights, and datasets all released, open-source video manipulation just took a serious step forward.
Alibaba's Qwen team just launched the Qwen3.5 Small Model Series, offering four compact sizes from 0.8B to 9B parameters along with base checkpoints. Built with native multimodal architecture and scaled reinforcement learning, the lineup aims to deliver snappy intelligence directly on edge devices and local agents.
ByteDance and Stanford researchers just open-sourced Spectrum, a training-free framework that accelerates image and video diffusion models by up to 4.7x. By forecasting future step trajectories with Chebyshev polynomials, it lets models like FLUX.1 and Hunyuan Video generate outputs much faster with virtually no drop in quality.
Google just launched gws, an open-source CLI covering the entire Workspace suite for both developers and AI agents. It dynamically parses APIs, returns structured JSON by default, and includes dozens of built-in agent skills so LLMs can run Gmail, Drive, and Sheets without custom wrappers.
Microsoft just released VibeVoice-ASR on Hugging Face, a speech-to-text model built to process up to 60 minutes of continuous audio in a single pass. By running across a 64K token window without slicing audio into chunks, it directly outputs timestamps, speaker diarization, and text across 50+ languages.
Pinchtab just launched a lightweight Go binary that lets AI agents control Chrome via a plain HTTP API. It features zero configuration, stealth anti-bot mode, and slashes token usage from 10,000 down to 800. Any agent framework or script can plug into it right away.
Researchers just open-sourced WildActor, a video generation framework that keeps an identity's face, body, and clothing consistent across unconstrained motions using only three reference images. Built on Wan2.2-5B, it supports both real people and cartoon characters while releasing the Actor-18M data construction pipeline.
SynapsCAD just launched an open-source tool aiming to become the Cursor for CAD. Developers write OpenSCAD code on the left and see real-time 3D geometry on the right, with the ability to click any part in the viewport to have AI modify the underlying code. Why it matters: Unlike black-box generative 3D meshes, every AI change remains clean, version-controlled code.
OpenAI just open-sourced Symphony, an orchestration tool that turns Linear boards into autonomous agent workflows. Instead of supervising coding models line by line, teams assign issues and let Symphony dispatch isolated Codex runs that implement features, pass CI, and record video walkthroughs for human sign-off. It shifts developer focus from watching agents code to managing the actual work.
OpenDataLoader PDF turns messy PDFs into clean Markdown and JSON at over 100 pages per second on a single CPU core, no GPU required. Built specifically for LLM and RAG pipelines, it preserves table structures, strips hidden prompt-injection text, and is completely free for commercial use.
PhysicEdit just open-sourced a physics-aware image editing framework that actually understands real-world causal dynamics. Built on top of Qwen ImageEdit, it reliably simulates complex visual physical phenomena like light refraction in water, melting butter, and collapsing card houses where most flagship models fail.
Nous Research just open-sourced Hermes Agent, an autonomous assistant that runs across Telegram, Discord, and your terminal. It features a built-in learning loop that turns completed tasks into reusable skills and builds a persistent model of you over time. It’s fully MIT-licensed and connects to any LLM provider.
Researchers just introduced VecGlypher, an open multimodal model that generates complete vector glyphs directly from text prompts or reference images. Instead of converting blurry raster graphics to vector paths, it autoregressively emits clean SVG outlines in one pass. It noticeably outperforms general models like GPT and Gemini in font stability, offering designers a powerful new foundation for custom typography.
Alibaba just dropped its Qwen 3.5 Medium series, featuring four models including Qwen3.5-Flash, 35B-A3B, 122B-A10B, and 27B. The team claims the 35B-A3B model outperforms the previous-generation 235B giant, proving architecture and RL can beat raw parameter scale. The Flash version also brings a default 1M context window and built-in agent tooling.
Alibaba just open-sourced CoPaw, a personal AI assistant featuring a web UI and multi-channel messaging support. It connects directly with DingTalk, Lark, Discord, and iMessage, running locally via Ollama or MLX with MCP modular tools. Basically, you get automated cross-channel workflows and daily summaries without writing a single line of code.
Researchers just open-sourced tttLRM, a novel 3D reconstruction model that uses Test-Time Training to convert multiple photos into crisp 3D Gaussian Splats. Unlike traditional 3DGS, it strips away foggy artifacts to capture fine wires and tiny text, all within a compact sub-4GB model that runs on consumer GPUs.
Guide Labs just open-sourced Steerling-8B, an 8B-parameter LLM built so every generated token can be traced back to its training data. By embedding an upfront "concept layer," the model hits roughly 90% of standard capability using less training data. Why it matters: It shifts model interpretability from fuzzy neuroscience into concrete engineering, making AI far easier to steer in regulated fields.
Someone just open-sourced Pascal Editor, a full 3D building editor that runs entirely inside your browser. Built with React Three Fiber and WebGPU, it lets users drag walls, stack levels, and reshape slabs in real time without paying for costly CAD or BIM licenses. Why it matters: It offers a free, web-native alternative to traditional desktop architecture software.
