New open-source and open-weight models like Llama, Qwen and DeepSeek, plus datasets, frameworks and developer tools worth trying.
Alibaba just open-sourced Wan2.1-FLF2V-14B, a video generation model designed to create smooth transitions between a user-provided first and last frame. It injects semantic features from both frames to keep visual style and complex motion consistent, and users can generate 5-second 720p clips for free on the official Wan website. The weights and code are now live on Hugging Face and GitHub.
Stability AI just teamed up with AMD to release ONNX-optimized versions of Stable Diffusion 3.5 and SDXL, engineered specifically for Radeon GPUs and Ryzen AI APUs. The new builds run up to 3.8x faster than standard PyTorch models while keeping full image quality intact. It is a major local AI performance win for AMD hardware users, ready to drop into ONNX Runtime workflows or Amuse 3.0.
Google just introduced Agent2Agent (A2A), an open protocol designed to let AI agents talk and collaborate across different platforms and vendors. Backed by over 50 partners including Atlassian and Salesforce, A2A complements Anthropic's MCP by focusing on multi-agent coordination. A production-ready release is slated for later this year.
Together AI and Agentica just open-sourced DeepCoder-14B-Preview, a lightweight coding reasoning model that punches way above its weight. Trained on 24K verified coding problems across 32 H100s, it scores 60.6% on LiveCodeBench to match OpenAI's o3-mini. The team also released the full dataset, training logs, and verl-pipe framework to democratize RL post-training for code.
San Francisco AI startup Deep Cogito just emerged from stealth and open-sourced Cogito v1, a lineup of hybrid reasoning models ranging from 3B to 70B parameters. Trained on top of Llama with an AlphaGo-style self-play loop, the models toggle between instant answers and deep reflection while adding native tool calling.
Meta just unveiled Llama 4 Scout and Maverick, marking its official pivot to open-weight, natively multimodal mixture-of-experts architectures. Both models operate on 17B active parameters, with Scout pushing context windows to a massive 10 million tokens, while an upcoming 2T-parameter teacher model, Behemoth, trains in the background.
Sentient Foundation just released Open Deep Search (ODS), an open-source framework designed to challenge closed AI search tools like Perplexity and ChatGPT Search. ODS pairs custom reasoning agents with web search, plugging into models like DeepSeek-R1 or GPT-4o. Paired with DeepSeek-R1, it even beat GPT-4o Search Preview on complex benchmark queries.
AudioX just open-sourced an all-in-one audio model that runs locally at under 6GB. It generates sound effects and background music from text, images, or silent video—even matching cuts to scene transitions—while handling seamless audio inpainting and continuation.
Perplexity just open-sourced pplx-kernels, a high-performance communication library for Mixture-of-Experts models that runs up to 10x faster than standard All-to-All primitives. It bypasses the CPU via GPU-to-NIC communication and overlaps data transfers with computation across NVLink, CX-7, and EFA networks. Why it matters: running massive MoE architectures across clusters just got substantially faster without sacrificing portability.
OpenAI just announced plans to release its first open language model since GPT-2 in the coming months, launching a community survey to guide development. CEO Sam Altman revealed the model will pack o3-mini-level reasoning and run directly on your own hardware, marking a major strategy shift amid fierce heat from open-source rivals.
OpenAI just announced that it is adding support for rival Anthropic’s open-source Model Context Protocol (MCP) across its entire product lineup. It is available immediately in the Agents SDK, with ChatGPT desktop and Responses API rollouts coming soon. Two fierce rivals are now aligned on the same data standard, making it far simpler to hook AI into real-world software.
Google just released TxGemma, an open family of models fine-tuned from Gemma 2 to predict molecular properties and analyze therapeutics data. Available in 2B, 9B, and 27B sizes, the models include specialized predict and chat variants that explain reasoning. The 27B predict model matches or beats previous SOTA on 64 out of 66 clinical and chemical tasks.
DeepSeek just quietly updated its flagship DeepSeek-V3 model, switching to a permissive MIT license that lets developers freely build commercial projects. Community benchmarks show a small boost in coding performance, and an Apple researcher managed to run a quantized version on a Mac Studio at 20 tokens per second. It's a huge win for open-source AI commercialization.
Open-source tool Browser Use just blew up after users realized it powers the viral AI agent Manus. Co-creator Gregor Zunic says daily downloads jumped from around 5,000 on March 3 to 28,000 on March 10. By letting models click buttons and fill out forms, it is rapidly becoming the shovel of choice in the web agent gold rush.
Google just released Gemma 3, its newest lineup of open-weight models ranging from 1B to 27B parameters. The 4B, 12B, and 27B versions handle both images and text with an expanded 128k token context window, supporting over 140 languages. Why it matters: The 27B model beats Gemini 1.5 Pro across key benchmarks, offering top-tier multimodal power you can run openly.
Reka just open-sourced Flash 3, a 21B parameter reasoning model built completely from scratch. Despite its lightweight footprint, Reka claims competitive performance and is already using it as the core engine behind its new enterprise intelligence platform, Nexus.
Tencent Hunyuan just open-sourced HunyuanVideo-I2V, releasing weights, inference code, and LoRA training scripts. The 13B parameter model turns a single image into a 5-second 2K clip complete with background audio, and even handles motion driving and lip-syncing. It is available on GitHub and HuggingFace, with a web demo open to the public.
LTX Studio just rolled out LTXV 0.9.5, an upgraded release of its open-source AI video model. The new version officially adds commercial licensing, keyframe conditioning, and native ComfyUI support, alongside higher resolution and fewer artifacts in longer video sequences.
IBM just open-sourced Granite 3.2 under the Apache 2.0 license, bringing experimental reasoning features and its first official vision language model. The new 2B and 8B Instruct models let developers toggle reasoning via a simple API parameter to save compute, while Granite Vision 3.2 2B targets enterprise document understanding.
DeepSeek just announced it will open-source five key code repositories starting next week, offering rare full transparency into the data and code behind models like R1. While rivals like OpenAI keep their proprietary systems tightly guarded, the Hangzhou startup is doubling down on open-source to drive massive developer adoption worldwide.
