DeepSeek is a Chinese AI research laboratory that has developed highly capable open-source language models including DeepSeek-V3 and DeepSeek-R1, notable for their efficiency and performance.
A new macOS app called T4B enables users to track token usage and AI subscription costs across eight major AI platforms—including Claude, OpenAI, Cursor, and Copilot—from a single menu bar interface. …
The MADSys Lab at Tsinghua University’s KTransformers project enables frontier-class MoE models like DeepSeek-R1 671B to run on commodity hardware with a CPU-GPU hybrid inference stack, achieving 286 …
Researchers at Doubleword developed UCCL-EP, an expert parallel communications kernel that enables GPU-initiated communication patterns like DeepEP to work across arbitrary NIC-accelerator pairs, incl…
MiniMax released the open-weights MiniMax M3 model on Modular Cloud, featuring a new Sparse Attention operation that achieves up to 15.6x speedup on decode while maintaining a 1 million token context …
Anthropic's Claude Fable 5 autonomously debugged a CSS bug across multiple browsers in a 15-step investigation, highlighting the rapid maturation of AI coding agents. OpenAI acquired Ona (formerly Git…
Local large language models have nearly caught up to frontier models for everyday tasks as of mid-2026, driven by engineering advances in sparse attention and mixture-of-experts architectures that red…
Hugging Face has released Open R1, an open-source reproduction of DeepSeek's R1 reasoning model, completing the first of three planned steps. The project includes a 350k-trace reasoning dataset distil…
Azure now offers DeepSeek V4 Pro and V4 Flash models through AI Gateway, providing users with an additional failover path for requests. The integration requires no code changes, as default routing aut…
A high-performance Expert Parallelism (EP) kernel is essential for running large Mixture-of-Experts (MoE) language models across multiple GPUs, as it handles the dynamic routing of tokens to experts l…
Self-hosted open models DeepSeek V4 Flash and Qwen 3.6 refuse to generate content about the 1989 Tiananmen Square protests and massacre, while willingly producing poems about other state violence like…
Lindy, an AI agent startup, switched its underlying model from Anthropic to DeepSeek and reports saving millions of dollars in inference costs. The company made the change to address the rising expens…
DeepSeek v4, an open-source model from a Chinese lab, achieved a 100x performance improvement on AMD hardware within 26 days of its Day 0 release, driven by the AMD SGLang engineering team. The Infere…
DeepSeek GUI launched as a desktop application that transforms the DeepSeek AI model from a browser-based chatbot into a local workspace for coding, document writing, project planning, and task automa…
AI Gateway data shows DeepSeek's share of tokens on the platform surged from under 1% to 17% in May, while its share of spend remained near 1%, as the low-cost model entered production use. Anthropic'…
A purported leaked screenshot of DeepSeek's graphical user interface shows an agent-first workspace layout that closely mirrors OpenAI's Codex interface, though the image remains unverified. The leak,…
A developer has launched a GitHub repository organized by stock ticker symbols, designed to preserve AI-generated financial research that is typically lost after a single query. The project allows any…
Three forces are reshaping AI cost structures as major buyers substitute cheaper models for expensive ones. Coinbase routed prompts to cheaper models to keep costs flat while token usage grew exponent…
A Huawei-led team in Shenzhen post-trained DeepSeek's 1.6-trillion-parameter model using a cluster of 1,000 Ascend 910C chips, according to the Shenzhen government. The achievement demonstrates China'…
A developer built a cost-saving solution combining Qwen-Turbo and DeepSeek series APIs, cutting total token costs up to 72% without reducing response quality. The system uses task-based model routing,…
A research group led by Huawei completed full-parameter post-training of DeepSeek's V4-Pro, a 1.6-trillion-parameter model, using a cluster of at least 1,000 Huawei Ascend 910C chips. The effort, whic…