New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Ai2 just released MolmoWeb, an open-weight visual web agent that operates purely from browser screenshots without parsing HTML or accessibility trees. Available in 4B and 8B sizes, it arrives with MolmoWebMix, a dataset of 30,000 human web trajectories, beating open-weight rivals on live benchmarks like WebVoyager.
ByteDance just open-sourced DeerFlow 2.0, an MIT-licensed agent orchestration harness designed for complex, multi-hour jobs. Instead of acting as a thin wrapper, a lead agent breaks tasks down and spins up isolated sub-agents inside Docker containers with dedicated filesystems and persistent memory. It works with local or cloud models like DeepSeek and GPT-4.
Pascal Editor just launched as a 100% free and open-source 3D architecture editor running entirely in the browser. Powered by WebGPU and React Three Fiber, it supports real-time editing of walls, slabs, and zones with GPU-accelerated rendering. It brings desktop-grade architectural modeling to the web without the typical enterprise software price tag.
GreenSock just rolled out gsap-skills, an official skill pack that teaches 40+ AI agents—including Cursor, Claude Code, and Copilot—how to write clean GSAP code. It covers timelines, ScrollTrigger, and React patterns, basically stopping your coding assistant from hallucinating outdated animation syntax.
MetaClaw just dropped an open-source continual meta-learning framework that turns regular conversations into training signals for LLM agents. It synthesizes reusable skills immediately after failures, then triggers cloud LoRA updates while you sleep or take meetings. That gives personal agents a way to self-improve over time with zero downtime and no local GPU required.
Unsloth just launched Unsloth Studio, an open-source local UI that turns PDFs or CSVs into datasets and fine-tunes models 2x faster with 70% less VRAM. It packages visual dataset creation, training, and GGUF export for 500+ open models into one interface, and is available now on GitHub.
SegviGen just open-sourced a 3D part segmentation framework built on TRELLIS.2 that splits meshes into individual components via single clicks or 2D guide images. It outperforms previous SOTA models by 40% in interactive segmentation while requiring far less supervision, though local inference demands at least 24GB of VRAM.
Google just open-sourced Spark VSSR, a powerful video super-resolution framework capable of sharpening degraded clips and restoring old films. The release includes inference models, full training code, and datasets, though running the 42.2GB package demands a high-end GPU. Basically, creators get state-of-the-art video upscaling that they can run and fine-tune locally.
Google just open-sourced the Colab MCP Server, enabling any MCP-compatible AI agent to programmatically control Google Colab notebooks straight from your terminal. Tools like Claude Code or Gemini CLI can now create cells, run Python scripts, and render visualizations live in the cloud. It basically turns Colab into an automated sandbox so agents no longer run risky, resource-heavy code on your local machine.
Researchers Albert Gu and Tri Dao just released Mamba-3, an open-source state space model architecture under the Apache 2.0 license designed to outpace Transformers. Its MIMO variant hits a 57.6% average benchmark accuracy—a roughly 4% relative bump over Transformers—while matching Mamba-2's quality using half the state size. By introducing complex-valued states and cutting GPU idle time, it targets reasoning gaps and slashes inference latency.
Mistral just unveiled Mistral Small 4, consolidating reasoning, multimodal vision, and agentic coding into a unified 119B MoE model with only 6B active parameters. It offers a 256k context window, a toggleable reasoning dial, and cuts completion latency by 40% compared to version 3. Best of all, it's released under Apache 2.0 and beats competing models with far fewer output tokens.
MiroMind just launched MiroThinker-1.7 and MiroThinker-H1, open-source heavy-duty research agents that bake real-time verification directly into their reasoning loops. Packing 235B parameters alongside a 61GB mini variant, H1 tops demanding benchmarks like BrowseComp by auditing intermediate steps and proof trails before serving answers.
Alibaba just open-sourced Fun-CineForge, a multimodal dubbing model that syncs speech directly to characters' lip movements and timestamps. By treating time as an explicit modality, it handles fast speaker switches and facial occlusions cleanly, turning automated film dubbing into a realistic studio workflow.
Tsinghua University researchers just open-sourced OpenMAIC, an AI platform that turns any topic or document into a full interactive classroom. Instead of watching static video lectures, users learn alongside AI teachers and classmates who speak, draw formulas on whiteboards, and engage in real-time debates. It basically shifts online learning into an active virtual classroom under a permissive MIT license.
NVIDIA just launched the Nemotron Coalition, teaming up with Mistral AI, Cursor, Perplexity, LangChain, and other top AI labs to build open frontier foundation models. Their first milestone is a joint base model co-developed with Mistral AI on DGX Cloud, which will be open-sourced to underpin the upcoming Nemotron 4 model family.
Tencent just open-sourced Covo-Audio-Chat, a 7B end-to-end audio language model. It processes continuous audio directly in a unified architecture to enable native, low-latency full-duplex conversations that can listen, speak, and handle interruptions on the fly. Weights and code are now available on Hugging Face and GitHub.
Texas A&M researchers just open-sourced Pulse of Motion, a lightweight AI designed to fix "chronometric hallucinations" in generated videos. Instead of guessing playback speed, the sub-2GB model calculates the true physical frame rate directly from visual dynamics and retimes the clip so movements actually obey gravity.
Fish Audio just launched Fish Audio S2, an open-source TTS model trained on over 10 million hours of audio across 80 languages. Instead of tweaking global mood sliders, users can insert plain-language bracket tags like [whispering] anywhere in their script for word-level control. Model weights and fine-tuning code are open on GitHub and Hugging Face.
Stanford researchers just launched OpenJarvis, an open-source framework designed to run personal AI agents entirely on your own devices. Their benchmarks reveal that on-device models can handle 88.7% of everyday chat and reasoning queries, effectively cutting reliance on cloud APIs and keeping private data where it belongs.
A security tracker called OpenClaw Exposure Watchboard just went live, revealing that over 1.07 million active OpenClaw deployments are sitting open on the public internet. More than 507,000 of them are in China and 312,000 in the US, with multiple instances already flagged for leaked credentials. Why it matters: Giving autonomous, high-privilege agents unrestricted public access without authentication is an open invitation for exploit.
