New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Researchers just open-sourced AnchorWeave, an interactive video generation framework that builds navigable 3D scenes from a single starting frame. Instead of hallucinating a new scene when you turn around, it uses retrieved local geometric memories to keep environments spatially consistent under keyboard control. Code is now available on GitHub.
rednote-hilab just released dots.ocr-1.5 on Hugging Face, a compact 3B multimodal OCR model. It claims SOTA multilingual document parsing across virtually any human script, with the standout trick being its ability to convert complex charts and diagrams directly into clean SVG.
ByteDance just open-sourced BitDance, an autoregressive image generation model that accelerates high-resolution synthesis by over 30x. By replacing codebook indices with binary visual tokens and decoding patches in parallel via continuous-space diffusion, it achieves a record 1.24 FID on ImageNet. Why it matters: AR models typically suffer from slow generation speeds, but BitDance proves they can rival diffusion in efficiency.
HKU Data Intelligence Lab just launched ClawWork, an evaluation benchmark that forces AI agents to earn their own living. Agents start with just $10 and must pay for their API token costs while completing real tasks across 44 professions—if they run out of cash, they go bankrupt. It basically replaces synthetic benchmark scores with raw economic viability.
Alibaba just dropped Mobile-Agent-v3.5 alongside the GUI-Owl-1.5 model family, spanning from 2B to 235B parameters. Built on Qwen3-VL, it natively unifies vision, reasoning, and tool execution across desktop, mobile, and browser environments, topping 20+ benchmarks. Model weights are now open on Hugging Face with live demos available.
ByteDance just open-sourced OpenViking under Apache 2.0, introducing a virtual file system to manage agent contexts, memories, and skills instead of relying on flat vector slices. It uses tiered loading across three summary levels to slash token overhead while fully visualizing the agent's retrieval path.
NVIDIA just dropped a guide on running the popular OpenClaw AI agent entirely for free on GeForce RTX GPUs and DGX Spark hardware. By pairing local runners like Ollama or LM Studio with open models, users can let an always-on agent manage local files and schedules without bleeding cash on cloud API tokens or handing over private data. It brings powerful autonomous desktop workflows right to consumer hardware.
Moonshine AI just released Moonshine Voice, an open-weight speech-to-text model engineered for real-time on-device use on iOS, Android, and Raspberry Pi. Its 200M parameter streaming model outperforms Whisper Large v3 in word error rate while featuring built-in voice activity detection and intent recognition.
Monarch RT just brought real-time AI video generation to consumer hardware, clocking 16 FPS on a single Nvidia RTX 5090. Its structured Monarch matrix attention runs up to 4x faster than standard FlashAttention while retaining about 95% of full-dense visual quality. Why it matters: developers no longer need multi-GPU server clusters to run live video diffusion models.
Alibaba just open-sourced DeepGen 1.0, a 5B unified model that handles both text-to-image generation and complex visual editing. Beyond modifying in-image text and shifting perspectives, it can even solve mazes, outperforming much larger models on editing benchmarks. The catch is its 72GB footprint on Hugging Face, putting consumer-grade hardware to the test.
NetEase Youdao just open-sourced LobsterAI, an MIT-licensed desktop agent for Windows and macOS. It natively edits Office documents, drives an in-app browser, and takes remote jobs via chat apps like WeChat and Telegram. It brings practical, approval-gated computer workflows directly to local desktops without complex terminal setups.
OpenMOSS just open-sourced the MOSS-TTS family, a versatile speech suite that clones voices from just a few seconds of reference audio. The lineup ranges from multi-speaker dialogue and voice design to ultra-low-latency streaming, with 8B and 1.7B checkpoints readily deployable on local GPUs.
Soul-AILab just open-sourced SoulX-Singer, a zero-shot singing voice synthesis model. Feed it a few seconds of raw speech and a reference melody, and it makes that voice sing whatever you want without fine-tuning. The best part: at under 3GB, it runs smoothly on basic GPUs and even CPUs.
ByteDance just open-sourced Protenix-v1, a biomolecular structure prediction model covering proteins, nucleic acids, and ligands. Under the exact same data cutoff and scale, it beats AlphaFold 3 across diverse benchmarks, releasing code, weights, and full data pipelines under Apache 2.0.
Shanghai AI Lab just open-sourced Intern-S1-Pro, a massive 1-trillion-parameter multimodal model tailored for scientific research. Built on a 512-expert MoE architecture that only activates 22B parameters per token, it tackles complex problems across chemistry, physics, and biology at Olympiad-level precision. It is free for commercial use, offering researchers a powerful open base for AI-driven discovery.
Open-source tool Just Dub It just released a lightweight LoRA adapter for the LTX-2 foundation model, letting you dub videos into languages like French and German while jointly generating matching lip movements. It preserves the speaker's facial identity and runs on consumer GPUs, with code and weights available on GitHub and Hugging Face.
UniAudio 2.0 just open-sourced a unified audio language model that handles text-to-speech, sound effects, music generation, and vocal style editing in a single architecture. Trained on 160B tokens, it proves multimodal audio no longer needs a stack of separate specialist models.
Open-source developers just launched PicoClaw, an ultra-lightweight AI assistant framework designed to replace bulky setups like OpenClaw. Built from scratch in Go, it uses under 10MB of RAM and boots in less than a second on $10 single-board computers. You can hook it up to Telegram or WhatsApp to handle tasks 24/7 without burning cash on hardware.
Mistral AI just dropped Voxtral Transcribe 2, introducing two speech-to-text models for batch and live audio. The standout is Voxtral Realtime, a 4B-parameter model released under Apache 2.0 that runs on-device with latency down to sub-200ms. It brings near-offline accuracy to real-time voice agents across 13 languages without sending private audio to the cloud.
See-Through just made 2D animation way faster by automatically splitting flat anime illustrations into up to 23 distinct, fully inpainted layers. It outputs a layered PSD complete with inferred depth maps, basically turning static art into ready-to-rig assets for Live2D.
