Total 121 · Clear filters
Runway published GWM Worlds 2 on 3 September 2026, a research preview that turns high-fidelity video and audio generation into a real-time interactive world model steered by text actions and camera motion.
On 8 July 2026 Mistral AI introduced Robostral Navigate, an 8-billion-parameter model that steers robots from camera images and spoken-style instructions, trained entirely in simulation and scoring 76.6% on a standard navigation benchmark.
On 9 April 2026 LG AI Research released EXAONE 4.5, its first open-weight multimodal model: a 33B vision-language system aimed at document understanding, with a 256K context window and the earlier EXAONE licence.
On 11 March 2026 Reka released Reka Edge, a 7-billion-parameter vision-language model built for physical AI on devices rather than in the cloud, with a 3.5-bit quantization format and comparisons against Qwen3.5 9B, Cosmos Reason2 8B and Gemini 3 Pro.
On 5 March 2026 Ai2 released Olmo Hybrid, a fully open 7B model that stacks Gated DeltaNet linear-RNN layers with attention layers in a 3:1 ratio and reaches Olmo 3's MMLU accuracy with 49% fewer training tokens.
TSMC's CoWoS advanced packaging places logic dies and HBM stacks on a silicon interposer; capacity remains sold out, and TSMC is expanding in Tainan and Chiayi while developing panel-level CoPoS.
Cambricon's next flagship cloud training chip, the MLU 690, entered mass production in early 2026, pairing a dual-die package with more than 700 TFLOPS of FP16 compute and 196 GB of HBM3 memory.
On 11 December 2025 Ai2 released Molmo 2, an open-weights family of vision-language models that adds pointing, tracking and grounded question answering in video, in 8B, 4B and 7B versions.
On 1 December 2025 Hugging Face released Transformers v5.0.0rc0, the first major version since v4 in November 2020. The library grew from about 40 architectures to more than 400.
On 19 November 2025 Yann LeCun confirmed that he would leave Meta after ten years to found an AI startup focused on advanced machine intelligence. Meta was to be a partner of the new company.
On 19 November 2025 Meta released SAM 3, which detects, segments and tracks objects from text or visual prompts, and SAM 3D, which reconstructs 3D objects and human bodies from a single image. Both were released as open models.
In November 2025 Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, an open-weight vision-language model that reasons with images — zooming into details and calling image search mid-thought — and reports a ChartQA score above Gemini 2.5 Pro and GPT-5.
On 20 October 2025 DeepSeek released DeepSeek-OCR under the MIT License. The model showed that rendering text as images can reduce the number of vision tokens needed for OCR, with decoding precision of about 97 percent at compression under 10x.
At Huawei Connect 2025 in Shanghai, Huawei published its first three-year Ascend chip roadmap, running to 2028, alongside new Atlas SuperPoD systems and a pledge to open source CANN.
ByteDance Seed posted the UI-TARS-2 technical report on 2 September 2025 and released it on GitHub on 4 September. It was trained with multi-turn reinforcement learning.
In September 2025 Alibaba released Qwen3-VL, a vision-language model family under Apache 2.0 whose flagship handles hours-long video, operates PC and mobile interfaces and does 3D grounding.
In September 2025 Alibaba's Qwen team released Qwen3-Next, an 80B-parameter open model with 3B parameters active per token, built on a hybrid of linear attention and sparse mixture-of-experts layers.
GLM-4.5V is a vision-language model released by Zhipu on 11 August 2025, built on GLM-4.5-Air and published under the MIT License; Zhipu reported state-of-the-art results among similarly sized models on 42 vision-language benchmarks.
On 31 July 2025 Cohere released Command A Vision, a 112-billion-parameter vision-language model built on Command A. Its open weights were posted under a non-commercial licence.
Aurora is a foundation model for the Earth system whose Nature paper, published on 21 May 2025 and promoted by Microsoft Research the same month, reports better than operational forecasts across four domains at far lower compute.
Physical Intelligence released π0.5 on 22 April 2025, a vision-language-action model built on its earlier π0 that the company says can carry out household tasks in homes it never saw during training.
