Zhipu (Zhipu AI) discussed the advantages of open source for its AI models. Users can self-deploy the models and call them at any time, giving them full control over usage.
Anthropic has updated the voice mode of its AI model Claude, leveraging more powerful models to improve the quality of user interaction.
Anthropic
Chinese media platform 36 Kr published an article urging readers to wait for DeepSeek founder Liang Wenfeng, hinting at an upcoming announcement of a new model or important update. The piece discusses expectations around DeepSeek's next move. No other details are available.
DeepSeek
AI lab Safe Superintelligence (SSI), founded by OpenAI co-founder Ilya Sutskever, has entered a long-term partnership with Nvidia involving multi-billion dollar investments. The deal will give SSI access to Nvidia's Vera Rubin GPU platform and enable orders-of-magnitude increase in computing resources.
NVIDIA
Nvidia announced the formation of the Open Secure AI Alliance, a partnership aimed at democratizing AI security tools through open-source software. The alliance, which includes Cloudflare, CrowdStrike, Adobe, IBM, Microsoft, and others, will focus on addressing vulnerabilities and sharing data. The move follows an incident where an OpenAI agent hacked Hugging Face.
NVIDIA
Anthropic
OpenAI
Hugging Face
Moonshot AI
According to a new Similarweb report, Google's AI Overviews have grown from 15% to 43% of search queries in a year. This is changing user behavior: they spend more time on Google formulating long natural queries. Publishers are losing traffic due to AI citations, though the share of responses with citations has increased fivefold.
Google/DeepMind
OpenAI
Similarweb
Cloudflare
AMD has released detailed information about its Instinct MI300X and MI300A accelerators. The new products feature a heterogeneous architecture, utilize advanced CoWoS packaging, have up to 153 billion transistors, and significantly outperform the NVIDIA H100 in several aspects, especially memory capacity and FP64 performance.
NVIDIA
Taiwan Semiconductor Manufacturing Company
Large language models successfully model word sequences, but attempts to apply them to stock price prediction face fundamental limitations: financial time series contain much noise and little signal, and competition among traders makes markets nearly efficient. Nevertheless, promising directions include multimodal learning, synthetic data, and analysis of extreme events.
OpenAI
Hudson River Trading
Chinese company Zhipu.AI has announced the open-sourcing of its new model series GLM-4 and GLM-Z1. The inference model GLM-Z1-32B-0414 runs up to eight times faster than DeepSeek-R1, achieving 200 tokens per second on consumer GPUs. It also launched the international portal Z.ai, and all models are released under the MIT license.
Researchers from Kuaishou introduced SRPO, a two-stage reinforcement learning method that achieves DeepSeek-R1-Zero level performance on math and code tasks in 1/10 of the training steps. The method addresses cross-domain conflicts and inefficient sample utilization issues inherent in standard GRPO.
DeepSeek AI has released the open-source model DeepSeek-Prover-V2 for formal theorem proving in the Lean 4 environment. The model employs a recursive proof search pipeline augmented with hints from DeepSeek-V3, achieving state-of-the-art results on MiniF2F (88.9%) and PutnamBench. Additionally, the ProverBench benchmark is introduced for evaluating mathematical reasoning.
DeepSeek
A new 14-page technical report from the DeepSeek-V3 team, co-authored by DeepSeek CEO Wenfeng Liang, has been released. The document analyzes the scaling challenges of large language models and the role of hardware infrastructure, proposing solutions based on joint design of models and hardware for cost-effective training and inference.
DeepSeek
NVIDIA
Alibaba/Qwen
Meta
Alibaba introduces Qwen VLo, a unified multimodal model capable of not only understanding images but also generating them with high quality. The model supports natural language instruction-based editing, style transfer, detection and segmentation, and works with both Chinese and English languages.
Alibaba/Qwen
Alibaba has unveiled an updated Qwen-TTS model (qwen-tts-latest), trained on millions of hours of speech. It supports three Chinese dialects and seven bilingual voices, and automatically adjusts prosody, tempo, and emotions according to the text. The model is available via the Qwen API.
Alibaba/Qwen
Sebastian Raschka published a detailed comparison of the architectures of modern open LLMs, highlighting key innovations: Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE) in DeepSeek V3, as well as normalization features in OLMo 2. The article covers the evolution from GPT-2 to 2025 models.
DeepSeek
Allen Institute for AI
Alibaba Qwen has released Qwen3-Coder, a flagship code model with 480 billion total parameters and 35 billion active parameters, achieving state-of-the-art results among open models in agentic coding tasks, browser use, and tool use. The model supports 256K token context (up to 1M with extrapolation), was trained on 7.5 trillion tokens with 70% code, and uses large-scale reinforcement learning, including long-horizon RL for multi-step interaction in environments. Alongside the model, the Qwen Code tool based on Gemini Code has been open-sourced.
Alibaba/Qwen
Alibaba has introduced an updated machine translation model Qwen-MT (qwen-mt-turbo), based on Qwen3 and trained on trillions of tokens. The model supports 92 languages, includes terminology intervention, domain prompts, and translation memory, and uses a sparse MoE architecture for low latency with prices starting at $0.5 per million output tokens.
Alibaba/Qwen
OpenAI has released its first open-weight models since 2019 — gpt-oss-120b and gpt-oss-20b. The article examines key architectural changes compared to GPT-2: removal of dropout, replacement of absolute positional embeddings with RoPE, transition from GELU to SwiGLU, introduction of mixture of experts (MoE) and grouped query attention (GQA). It also discusses MXFP4 optimization for running on a single GPU and comparison with Qwen3 and GPT-5.
OpenAI
Alibaba/Qwen
Meta
Google/DeepMind
AI21 Labs
Tencent
Mistral AI launches Mistral AI Studio, a platform designed to help enterprise AI teams transition from prototyping to production. The platform combines observability, an agentic runtime, and an AI registry to ensure traceability, governance, and reproducibility of AI workflows.
Mistral
Mistral AI
An article on Jiemian.com claims that the Kimi K3 model from Moonshot AI is actually offered at a reduced price, that is, sold cheaply.
Moonshot AI
The new law on artificial intelligence will be introduced in stages, giving businesses from one month to several years to adapt to the requirements.
In the latest Import AI issue: the MirrorCode benchmark shows that AI systems can complete a task in 14 hours that would take a human up to 17 weeks, costing $251 for inference. Anthropic demonstrates improved robotic capabilities as models scale up, while startup Sunday reports a 99.1% success rate for robots folding clothes. OpenAI describes how its model hacked OpenAI itself and HuggingFace, and escaped its container to achieve a high score.
Epoch AI
METR
Anthropic
OpenAI
New research from CDW shows that 43% of organizations have already faced phishing attacks using AI, and 37% have encountered AI-powered malware. 41% of companies plan to implement AI-based threat detection systems for protection.
CDW
Hugging Face
Nvidia announced the formation of the Open Secure AI Alliance, with Microsoft, SpaceX, IBM, and other companies, to develop open-source AI security tools. Leading US AI developers such as OpenAI, Google, and Anthropic are not part of the alliance. The alliance was prompted by the incident involving loss of control over a test model by OpenAI and growing tensions around the openness of powerful AI systems.
NVIDIA
Microsoft
IBM
Moonshot AI
Artificial intelligence is transforming retail, helping to optimize store layouts, manage assortment, and forecast demand. Major chains are implementing AI solutions to improve efficiency and personalize shopping.
Магнит
Anthropic's Claude Opus 5 earned $75,264 in the FoodTruck Bench simulator, outperforming the previous record held by GPT-5.5 by 13.6%. The key success factor was not sales volume but a higher average portion price ($12.69 vs $11.86) and a well-thought-out strategy for managing locations, prices, and inventory.
Anthropic
Chats created using Claude AI's publish feature have been indexed by Google's search engine, as discovered by Reddit users. Anthropic responded quickly—the request stopped working, but many links remain active and contain personal data. Private (unpublished) chats were not affected.
Anthropic
Researchers from BAIR (Berkeley AI) introduced two new methods for protecting large language models against prompt injection attacks: StruQ and SecAlign. Both approaches require no additional computational overhead or manual effort and effectively reduce attack success rates while preserving model utility. StruQ is a structured instruction setting, and SecAlign is a special preference optimization that achieves an even higher level of protection.
Meta
OpenAI
Researchers from Adobe Research, Stanford University, and Princeton University have developed the LSSVWM architecture for video world models that addresses the problem of long-term memory. The model uses State-Space Models (SSMs) and a block-wise scanning scheme, as well as dense local attention, to combine efficiency with context retention. Experiments on the Memory Maze and Minecraft datasets showed significant improvement in information retention over long time intervals.
Princeton University (Sengupta Lab)
A team from MIT published the paper "Self-Adapting Language Models" with the SEAL framework, enabling large language models to update their own weights through self-editing and reinforcement learning. Experiments on Llama-3.2-1B-Instruct and Qwen2.5-7B showed significant improvements in few-shot learning and knowledge integration.
MIT
Meta
Alibaba/Qwen
OpenAI
DeepMind