New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Alibaba just released Qwen3-Coder-Next, an open-weight model built specifically for coding agents and local developers. It uses agentic training across 800K verifiable tasks and packs an 80B total parameter base into just 3B active params, delivering strong SWE-Bench Pro results while integrating smoothly with tools like Cline.
OpenBMB just rolled out MiniCPM-o 4.5, a 9B omnimodal model that sees, listens, and speaks simultaneously in real time. Thanks to an end-to-end full-duplex design, it doesn't just wait its turn—it monitors inputs at 1Hz and can proactively interrupt with reminders. It matches Gemini 2.5 Flash on key multimodal tasks while running locally on a Mac or single GPU.
Zhipu AI just open-sourced GLM-OCR, an ultra-compact 0.9B multimodal model built for heavy-duty document understanding. Despite its small size, it topped OmniDocBench V1.5 with a 94.62 score, parsing intricate tables, formulas, and handwriting into Markdown or structured JSON. It runs out of the box on vLLM and Ollama under an MIT license, making local document extraction fast and dirt-cheap.
Omnimat Zero just open-sourced a video tool that erases objects along with their tricky reflections and ground shadows. It can also isolate foreground subjects into transparent layers for background swaps, making clean video cleanup far less tedious.
ACE Studio and StepFun just open-sourced ACE-Step 1.5, a commercial-ready music foundation model built for consumer hardware. It generates full songs in under 10 seconds on an RTX 3090, supports 50+ languages, and runs with less than 4GB of VRAM. It basically puts capable music generation directly on your desktop under an MIT license.
MOSS-Audio just open-sourced a unified audio-language model available in 4B and 8B sizes, each offering Instruct and Thinking variants. It packs speech recognition, speaker analysis, emotion detection, environmental sound, and music understanding into one model, significantly outpacing Gemini-3.1-Pro on timestamped ASR accuracy.
Anthropic just open-sourced Knowledge Work Plugins, a collection of role-specific workflows for sales, legal, marketing, and more. Built entirely on Markdown and JSON with zero code, each plugin equips Claude Cowork and Claude Code with domain playbooks and tool connectors. Teams can fully customize them simply by editing text files.
WordPress contributors just released wp-playground, a dedicated skill that lets AI agents spin up isolated WordPress test sites via the Playground CLI in seconds. The tool automatically detects whether generated code is a plugin or theme, mounts it, and lets agents like Claude Code verify behavior using tools like Playwright. It closes the feedback loop so AI can build and debug without manual setup.
OpenMOSS just released MOVA, an open-source model that generates high-fidelity video and synced audio in a single pass. It runs on a 32B MoE architecture with 18B active parameters, handling precise lip-sync without stitching separate models together. Weights are live on GitHub and Hugging Face, though the 720p version requires serious GPU muscle.
robbyant just open-sourced LingBot-World, a real-time interactive world model built on Wan2.2. It generates controllable 480P and 720P environments from text or images at 16 FPS with sub-second latency, while retaining consistent scene memory over minutes. Why it matters: it gives developers a practical open alternative for interactive worlds and robot learning.
Researchers just released TeleStyle, an open-source model that restyles both images and videos using reference pictures or text prompts. Built on Qwen-Image-Edit, it pairs curriculum learning with a dedicated video-to-video module to keep content and motion consistent across frames.
Alibaba's Tongyi Lab just open-sourced Z-Image, the undistilled foundation model behind its fast Turbo edition. Built specifically for customization, it fixes the Turbo version's repetitive faces and compositions while adding negative prompt support. It needs 50 steps instead of 28, and ComfyUI already supports it out of the box.
Clawdbot (now Moltbot) just went viral as an open-source, self-hosted AI assistant that lets you control your local computer through Telegram or WhatsApp. It can run scripts, manage files, and even write its own code to build new skills without sending private data to the cloud.
Letta just launched Claude Subconscious, an open-source plugin that runs a secondary agent underneath Claude Code. It quietly studies your repo, stores context across 8 persistent memory blocks, and whispers guidance back into prompts without blocking your workflow.
Nvidia just launched the Earth-2 family of open models, delivering the world's first fully open, accelerated software stack for AI weather and climate forecasting. The suite lets organizations run and fine-tune models on their own hardware, cutting compute time by up to 90% in operational tests by meteorological agencies. Why it matters: It shifts heavy supercomputer weather forecasting into accessible AI workflows that any enterprise or government can run locally.
Open-source local voice tool Voicebox combines instant voice cloning with system-wide dictation, racking up over 42,000 GitHub stars. It can clone voices from a few seconds of audio, type into any text field via global hotkeys, and voice coding agents like Claude Code. Basically, it lets you ditch paid voice subscriptions and run your entire workflow locally.
Microsoft just open-sourced VibeVoice-ASR, a 9B speech model capable of transcribing up to 60 minutes of continuous audio in a single pass. Instead of slicing audio into chunks, it unifies transcription, speaker diarization, and timestamps into one end-to-end generation task. It basically solves context fragmentation for long meetings and podcasts.
Tencent just released HunyuanImage 3.0 Instruct, an open-weights image-to-image model that uses chain-of-thought reasoning to plan edit steps and preserve untouched regions before rendering. It handles lighting, perspective shifts, and multi-image merging, though self-hosting its 169 GB weights requires massive VRAM.
Researchers just released RF-DETR segmentation under an Apache 2.0 license, introducing a new real-time vision baseline. It arrives in six model sizes, ranging from a Nano variant clocking 3.4 ms per image (40.3 mAP) to a 2XLarge variant reaching 49.9 mAP at 21.8 ms. Why it matters: Developers get an unconstrained, ultra-fast segmentation backbone built for edge devices and latency-critical vision tasks.
Alibaba just released Qwen3-TTS, an open-source speech model that clones any voice from just seconds of audio and lets you design new vocal styles using text prompts. Packing a dual-track architecture with 97ms end-to-end latency, it can fluidly generate cross-lingual speech and regional dialects for live streaming and AI agents.
