Stanford researchers just released Tokasaurus, an open-source LLM inference engine built specifically for high-throughput batch workloads like synthetic data generation. By slashing CPU overhead, dynamically grouping shared prefixes, and optimizing pipeline parallelism for non-NVLink GPUs, it beats vLLM and SGLang by over 3x on key benchmarks.
Tencent Hunyuan and Tencent Music just open-sourced HunyuanVideo-Avatar, a new model that brings static photos to life from audio clips. It automatically reads scene context and emotion to generate lifelike speech or singing videos up to 14 seconds long. The single-character mode is out now, with multi-character interactions coming soon.
WordPress just created an official AI team with contributors from Google and Automattic to drive platform-wide artificial intelligence projects. Taking a "plugin-first" playbook from its performance team, the group plans to iterate quickly with canonical plugins without waiting for slow Core release cycles. Why it matters: It keeps the web's biggest CMS unified and open as AI shifts online creation.
Mistral just unveiled Devstral, a 24B coding model built with All Hands AI under a fully commercial Apache 2.0 license. It fits on a single RTX 4090 or a 32GB Mac, beating Google's Gemma 3 27B and DeepSeek's V3 on SWE-Bench Verified. Why it matters: Developers finally get a capable, agent-ready code model they can run locally without restrictive licenses.
Stability AI just released Stable Video 4D 2.0, an upgraded diffusion model that turns a single object-centric video into dynamic 4D assets and novel views without multi-camera setups or reference images. Weights and code are available under its Community License on Hugging Face and GitHub.
Google just unveiled Gemma 3n, a mobile-first open model designed to run rich multimodal AI entirely on phones and laptops. Thanks to a new Per-Layer Embeddings technique, its 5B and 8B models operate with dynamic memory footprints of just 2GB and 3GB. Why it matters: it makes complex, private local AI actually fit on everyday hardware.
Stability AI just teamed up with Arm to open-source Stable Audio Open Small, a 341M parameter text-to-audio model built to run on smartphone CPUs. Powered by Arm KleidiAI, it generates up to 11 seconds of stereo sound in under 8 seconds. Why it matters: generative audio is finally moving off the cloud and directly onto edge devices for real-time sound effects.
Lightricks just launched LTX Video-13B, a 13-billion-parameter open-source video model available on GitHub and Hugging Face. It brings multi-scale rendering that generates high-res clips up to 30 times faster than peers, and clever quantization lets it run on consumer GPUs. Plus, it's free to license commercially for teams under $10M in revenue.
Nvidia just launched Parakeet-TDT-0.6B-v2, an open-source speech recognition model that can transcribe an hour of audio in just one second on GPU hardware. The 600M-parameter model tops the Hugging Face Open ASR leaderboard with a 6.05% word error rate, approaching proprietary rivals. Best of all, it is released under a permissive CC-BY-4.0 commercial license.
Ai2 just released Olmo 2 1B, a 1-billion-parameter open model that outperforms Google's Gemma 3 1B and Meta's Llama 3.2 1B on math reasoning and truthfulness benchmarks. It comes with complete code and training datasets under Apache 2.0, running easily on consumer laptops and mobile devices.
Microsoft just launched Phi-4-reasoning-plus, an open-weight 14B model that punches way above its weight. Trained on curated synthetic data and GRPO reinforcement learning, it outperforms DeepSeek-R1-Distill-70B on the AIME 2025 math benchmark. Even better, it drops under a permissive MIT license and supports Ollama and vLLM right out of the box.
Intel just detailed AutoRound, a post-training quantization tool designed for LLMs and vision-language models. By optimizing weight rounding and clipping ranges together, it boosts relative accuracy at INT2 by up to 2.1x over popular baselines. Quantizing a 72B model takes just 37 minutes, with direct exports to GPTQ, AWQ, and GGUF.
OpenAI is reportedly planning a "handoff" feature for its upcoming open reasoning model, letting it ping larger cloud models when local queries get too tricky. The model is being trained from scratch for an early summer launch to challenge DeepSeek R1. It is a clever move that bridges open weights with OpenAI's paid API ecosystem.
Korean startup Nari Labs just released Dia, an open-source 1.6B text-to-speech model built by two undergraduates with zero outside funding. The model handles emotional tones, multi-speaker tags, and nonverbal cues like laughter and coughing, outshining commercial leaders like ElevenLabs in benchmark tests. It proves solo builders can now challenge billion-dollar labs.
Google just released optimized versions of its Gemma 3 open models using Quantization-Aware Training (QAT). By simulating low-precision operations during training, the 4-bit (int4) weights cut memory demands drastically while slashing perplexity loss by 54%. This means developers can now run powerful 27B models locally on consumer cards like the NVIDIA RTX 3090 instead of expensive cloud clusters.
LTX Studio just released LTXV 0.9.6, introducing two 2B parameter open-source video models. Creators can use a full-quality variant for polished final outputs, or a lighter distilled model built for fast iteration and easier local deployment.
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.
