Auto Added by WPeMatico
Z.ai just released GLM-5.3. GLM-5.3 runs on the same 743B base model as GLM-5.2. Every reported gain comes from scaled post-training: more task environments, more environment types, longer training. The results land in two places. Coding jumps most on the longest-horizon benchmarks, with Terminal-Bench 3.0 moving from 4.6 to 28.3. Cybersecurity moved further than Z.ai […]
Cactus Compute has released Needle 2, an open 45M-parameter model for tool calling, device use, and structured extraction. The entire model ships as a single 14MB binary that runs a full session in about 28MB of RAM. Weights are trained and deployed at CQ2-bit using Cactus Quants, and the model is sealed inside the company’s
Google has released Gemini 3.7 Flash, the newest model in its Flash tier, three weeks after Gemini 3.6 Flash. The model card describes it as a refinement of 3.6 Flash with algorithmic improvements to the core reasoning foundation — not a new pretraining run. It accepts text, images, audio, and video across a 1M-token context
Yesterday, Liquid AI released LFM2.5-VL-3B. It is a 3.1B-parameter vision-language model built for on-device deployment. The model reads digital screens across mobile, web, and desktop. It grounds objects to coordinates, parses documents and charts, and calls tools from text or image input. Liquid AI reports an average of 69.4 across 28 vision benchmarks. That matches
Dyna Robotics has released Dyna-2, a world-action model for robot manipulation. It was pre-trained on more than one million hours of egocentric human video. That is roughly 170 years of continuous waking experience. Robot learning has been bottlenecked by action-labelled data, which teleoperation must deliberately produce. Dyna-2 tests whether ordinary human video can substitute. The
SpaceXAI just released Grok 4.6. The release is a post-training upgrade over Grok 4.5 rather than a larger base model. SpaceXAI held the foundation constant and spent the improvement on a longer supplemental training run, regenerated supervised fine-tuning trajectories, and reinforcement learning in agentic environments. Agents that stay on a task across many steps without
NVIDIA introduced open technologies for building always-on AI agents from systems of specialized models. Two artifacts shipped together. Nemotron 3.5 Lightning is a lightweight, customizable open model built for high-volume agentic tasks, and NeMo Switchyard is an open source routing library that directs each step of an agent workflow to the most capable and efficient
Object removal models have improved faster than the metrics used to judge them. Diffusion erasers now reconstruct shadows, reflections and occluded structure convincingly, yet PSNR, SSIM, LPIPS, ReMOVE and CFD frequently rank their outputs the wrong way. The root cause is structural: erasure is an ill-posed, one-to-many task, so no single ground truth exists to
webAI has released TwIL-LM, a two-model family of formal-logic reasoners at 1.7B and 3B parameters. The 3B member, TwIL-LM3, is a merged fine-tune of SmolLM3-3B; the 1.7B member is a PEFT LoRA adapter for SmolLM2-1.7B-Instruct. Both target autoformalization: translating English into first-order logic and checking whether a conclusion follows from its premises. Both run locally,
Meta has released Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark. It is tuned for always-on local agent workflows, and ships under Apache 2.0. A 30B model normally needs over 55 GB of memory at full precision. Meta compresses it to roughly 4-bit, then adds block-level speculative decoding so it answers fast enough
