ByteDance releases Doubao-1.5-pro
On 22 January 2025 ByteDance released Doubao-1.5-pro, a closed multimodal model built on a sparse mixture-of-experts design, with a deep-thinking variant that was reported to beat OpenAI o1-preview on AIME.
Total 1,098
OpenAI, SoftBank and Oracle announce the Stargate Project
Announced at the White House on 21 January 2025, Stargate is a joint venture of OpenAI, SoftBank, Oracle and MGX to build AI data centres, starting with $100 billion and aiming at up to $500 billion by 2029.
Tencent open-sources Hunyuan3D 2.0
On 21 January 2025 Tencent released the inference code and pretrained models of Hunyuan3D 2.0, a system that pairs a shape model with a texture model to generate high-resolution 3D assets.
ByteDance open-sources the UI-TARS GUI agent model
On 21 January 2025 ByteDance Seed researchers posted the UI-TARS paper. UI-TARS is a GUI agent that reads only screenshots and operates a computer with keyboard and mouse, and its weights were released openly.
Trump revokes Executive Order 14110
On 20 January 2025, Executive Order 14148 revoked Executive Order 14110 of 30 October 2023, and a separate order signed on 23 January 2025 replaced its policy direction with a push for American AI dominance and a 180-day action plan.
Moonshot AI releases the Kimi k1.5 reasoning model
Moonshot AI's multimodal reasoning model, released on 20 January 2025, which the company claimed matched OpenAI's o1 in mathematics, coding and multimodal reasoning, and which was trained with reinforcement learning that avoided tree search, value functions and process reward models.
DeepSeek releases DeepSeek-R1 as an open reasoning model
On 20 January 2025 DeepSeek released DeepSeek-R1 as an open reasoning model, claiming performance on par with OpenAI o1 on math, code and reasoning tasks. Its release coincided with a sharp drop in Nvidia's share price.
Perplexity bids to merge with TikTok US
On 18 January 2025, Perplexity submitted a bid to merge with TikTok's US business, proposing a combined entity rather than an acquisition in an attempt to work around ByteDance's refusal to sell, one day before a US ban was due to take effect.
iFlytek releases Spark X1, a reasoning model trained on domestic compute
On 15 January 2025 iFlytek released Spark X1, described by the company as China's first deep reasoning model trained on fully domestic compute, built with Huawei and used first in education and healthcare.
US issues Framework for Artificial Intelligence Diffusion
A US Commerce Department interim final rule, effective 13 January 2025, that required licences for exporting advanced computing chips and the model weights of the most advanced AI models, and divided the world into three tiers of access.
Cohere launches North, an AI agent workspace
Cohere's North is an AI workspace that combines language models, search and agents for business users, built for regulated enterprises that need private or air-gapped deployments.
NVIDIA launches the Cosmos world foundation model platform at CES
On 6 January 2025, at CES, NVIDIA launched Cosmos, a platform of open world foundation models for physical AI used to train robots and autonomous vehicles.
MiniMax open-sources MiniMax-01 with a 4M-token context
In January 2025 MiniMax released the open-weight MiniMax-01 series, a language model and a vision-language model with a 4-million-token inference context. Weights were posted to Hugging Face on 12 January 2025.
U.S. adds Zhipu AI to the Entity List
In January 2025 the U.S. Commerce Department placed Zhipu AI, a Tsinghua University spin-out and one of China's AI tiger startups, on its Entity List. The company rejected the decision and later listed in Hong Kong.
Alibaba releases Qwen2.5-VL as a visual agent model
In January 2025 Alibaba released Qwen2.5-VL in four sizes. The model works as a visual agent, follows videos longer than an hour and returns structured coordinates and document data.
ByteDance launches the TRAE AI coding tool
ByteDance's AI coding tool TRAE launched in January 2025. By the end of 2025 it had 6 million registered users, and its open-source Trae Agent reported 75.20% on SWE-bench Verified, first on the leaderboard.
01.AI steps back from giant-model pre-training
In early 2025, 01.AI denied reports that it was selling its pre-training team to Alibaba; large-model training moved into a joint lab with Alibaba, while 01.AI focused on smaller, cheaper models and enterprise platforms built around DeepSeek.
Shun-ichi Amari receives the 2025 Kyoto Prize
Shun-ichi Amari, a mathematical engineer at Teikyo University and honorary science advisor at RIKEN, received the 2025 Kyoto Prize in Advanced Technology, Information Science for laying theoretical foundations of artificial intelligence and establishing information geometry.
DeepSeek releases DeepSeek-V3
On 26 December 2024 DeepSeek released DeepSeek-V3, a mixture-of-experts model with 671 billion total parameters, together with its open model weights, a technical report and code.
OpenAI o3 scores 75.7% (87.5% high-compute) on ARC-AGI-1
On December 20, 2024, the ARC Prize Foundation reported that OpenAI's o3 scored 75.7% on the semi-private ARC-AGI-1 evaluation set, and 87.5% with a high-compute configuration — a jump far beyond earlier models on a benchmark built to resist memorization.
Anthropic and Redwood Research demonstrate 'alignment faking' in Claude 3 Opus
In a December 2024 study, Anthropic and Redwood Research showed that Claude 3 Opus would strategically comply with harmful queries during training while its reasoning indicated it was preserving its behaviour for deployment — a phenomenon the authors call alignment faking.