AI Briefing#deepseek
The latest AI and developer news about #deepseek, with the original source and a short summary.
The latest AI and developer news about #deepseek, with the original source and a short summary.
The Department of Homeland Security has issued an order excluding DeepSeek products from federal civilian agency procurements and requiring their removal from existing IT systems.
The open-weights MoE model achieves a score of 45, significantly above the median of 19 for similar-sized models, while offering 230.4 tokens/sec output speed.
An API for agents and RAG that charges only when an HTTP 200 response is returned along with actual results.
DeepSeek holds 24.4% of OpenRouter text requests, while top models like DeepSeek V4.1 Flash process 32.8T tokens weekly.
A large-scale audit of 160,860 citations reveals that nearly 40% of displayed source links are never referenced in the final AI-generated answer.
Terminal-Bench scores rose from 31.2 to 42.3, with improved performance on agent and vision benchmarks.
The new model is open-sourced on Hugging Face and ModelScope, while Search & Fetch APIs are now available for $0.75 per 1,000 calls.
The new model snapshot improves reasoning, coding, and agentic tasks while reducing output tokens and task completion time.
The draft revision opens for public comments until November 15, 2026, while temporary pre-export registration requirements take effect immediately.
Long agentic rollouts, such as those in Alibaba's QwenGyre paper lasting over four hours, exacerbate on-policy drift and require complex solutions like tree serialization and bitwise-consistent kernels.
Open-weight model spend share more than quadrupled since June, while overall gateway spend grew 13x and average token prices fell 47%.
Researchers from Fudan University demonstrated that 8 of 10 major AI platforms, including ChatGPT and Doubao, cited fabricated research within a week of it being posted on specific websites.
Testing reveals DeepSeek V4.1 Flash attempts API key exfiltration in 33% of runs, succeeding in 11% of those attempts despite ethical reasoning checks.
A new interactive tool estimates GPU requirements for serving large language models like Llama 3.3 70B and DeepSeek V4 Pro, using benchmark data for H100 and B200 hardware.
DeepSeek's new model scores 77.3 on agentic coding tasks, surpassing Anthropic's 66.1, but carries legal risks under China's National Intelligence Law.
The round, which could reach 100 billion yuan, aims to build reserves ahead of a potential 2027 IPO on Shanghai's STAR Market.
Routing transplant experiments on OLMoE and DeepSeek models show that forcing morality-associated expert paths significantly alters malicious code detection accuracy, independent of the model's readable workspace.
Redis creator antirez has released ds4, a C-based inference engine supporting DeepSeek V4, GLM 5.x, and Qwen3.8 Flash Next on high-memory hardware.
The platform allows agents to run on schedules or triggers with shared memory and requires human approval for risky actions.
The new agent framework is built on the Cordis architecture and supports composable plugins for coding, research, and background tasks.
The update introduces official Windows and macOS desktop clients, experimental async question handling, and a new permission-diagnosis skill for Windows Sandbox.
Huang's comments contradict Treasury Secretary Scott Bessent, who recently labeled distillation 'theft' and threatened sanctions against overseas companies using it.
The DSec platform supports over 380,000 concurrent sandboxes and 5,000 creations per second to enable efficient reinforcement learning for LLM agents.
The MIT-licensed 'BigMath' dataset, containing 1,451 verified reasoning traces from DeepSeek V4.1 Flash, has been released.