Total 24 · Clear filters
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.
In June 2025 EleutherAI and its partners released Common Pile v0.1, an 8 TB collection of public-domain and openly licensed text for training language models. The team also used it to train two 7B models.
NVIDIA announced Blackwell Ultra at GTC in San Jose on 18 March 2025, a platform aimed at AI training and at inference that spends more compute per answer, alongside new HGX B300 NVL16 systems and named server and cloud partners.
On 13 March 2025, Ai2 released OLMo 2 32B, which it called the first fully open model to beat GPT-3.5-Turbo and GPT-4o mini on a suite of academic benchmarks, and published everything needed to reproduce it.
On 6 March 2025 Mistral AI launched Mistral OCR, a document-understanding API that it said processes up to 2,000 pages per minute, handles thousands of scripts and languages, and scored above GPT-4o on its benchmark.
On 26 November 2024 Ai2 released OLMo 2 in 7B and 13B sizes, publishing data, code, recipes and checkpoints alongside the weights, and described the pair as the best fully open language models to date.
In November 2024 Hugging Face released SmolLM2, a family of small language models at 135M, 360M and 1.7B parameters, with open weights, an instruct variant and a technical report that made the recipe public.
On 16 July 2024 Hugging Face released SmolLM, a family of small models for on-device use in 135M, 360M and 1.7B sizes, together with SmolLM-Corpus, the training dataset behind them.
In July 2024 SoftBank Group acquired Graphcore, which became a wholly owned subsidiary and kept its name and offices; the deal was recorded at about US$500 million, subject to UK government review.
On 14 June 2024 NVIDIA released Nemotron-4 340B, a family of base, instruct and reward models built for generating synthetic training data. It published the models under a permissive license, and the reward model topped RewardBench.
On 28 February 2024 BigCode and NVIDIA released StarCoder2 in 3B, 7B and 15B sizes. The new training set, The Stack v2, is 67.5 TB and is built from the Software Heritage archive.
On 1 February 2024, the Allen Institute for AI released OLMo, a language model released together with its pretraining data, training code, logs and more than 500 checkpoints per model, all under the Apache 2.0 licence.
In February 2024 Cohere For AI released Aya, an instruction-tuned model covering 101 languages. It has 13 billion parameters, uses an Apache 2.0 licence and beats mT0 and BLOOMZ on most tasks, according to its authors.
On 13 November 2023 NVIDIA announced the HGX H200 at the SC23 conference in Denver, an addition to its Hopper line that twelve server partners and four major cloud providers were named to offer, with shipments expected from the second quarter of 2024.
On 11 September 2023 Microsoft researchers posted the phi-1.5 technical report, a 1.3-billion-parameter model for natural-language common-sense reasoning, released openly under the MIT licence.
On 20 June 2023 Microsoft Research posted the paper 'Textbooks Are All You Need' to arXiv, introducing phi-1, a 1.3-billion-parameter code model trained on textbook-quality data.
On 5 April 2023 Meta released the Segment Anything Model (SAM), a promptable image segmentation model, together with the SA-1B dataset of over 1.1 billion masks. Meta presented it as a step toward a foundation model for image segmentation.
On 11 February 2019 the US president signed an executive order creating the American AI Initiative, a coordinated federal strategy that made AI a budget priority and set agencies deadlines on data access, standards and regulation.
On 16 May 2012 Google announced the Knowledge Graph, an attempt to move search from matching strings to identifying entities; at launch it held more than 500 million objects.
The 2011 ILSVRC saw XRCE, the 2010 runner-up, win classification with compressed high-dimensional signatures and linear SVMs, while the new single-object localization task was won by UvA's selective search approach.
The spring 2010 release of ImageNet, announced by Fei-Fei Li on 11 May 2010, totalled more than 11 million images in 15,500 categories and added bounding boxes, SIFT features and a MATLAB toolbox, alongside a call to join ILSVRC2010.
