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 technical AI safety researcher argues that safety-focused researchers should engage with capabilities research, highlighting On-Policy Self-Distillation (OPSD) as a promising alternative to GRPO for…
A developer's analysis shows that the same DeepSeek V4 Flash model produces different long-task outcomes depending on the agent runtime, such as Codex versus Claude Code. The developer argues that the…
ThinkAny AI released DScode, an open-source coding agent powered by DeepSeek, available via npm as @thinkany/dscode. The tool runs locally in any repository, supports parallel agents, sandboxed comman…
DeepSeek's V4-Flash-0731, a 284-billion-parameter Mixture-of-Experts model with about 13B active parameters per token, scored 50 on Artificial Analysis's Intelligence Index, one point below GLM-5.2 an…
DeepSeek founder Liang Wenfeng, in a leaked 3-hour-44-minute investor meeting transcript, revealed the lab's 'Costco strategy' for achieving AGI: pricing its API so GPU hardware pays for itself in ten…
DeepSeek released V4 Flash, a coding model that performs near Anthropic's Claude Opus 4.8 but costs about 28 cents per output unit versus $25 on Opus 4.8, a 99% discount, accelerating the commoditizat…
DeepSeek founder Liang Wenfeng said his AI startup, valued at $60 billion, operates without KPIs or overtime culture, a philosophy revealed in a nearly four-hour staff discussion transcribed by Tencen…
DeepSeek founder Liang Wenfeng says his AI startup, valued at $60 billion, operates without KPIs or overtime, telling staff in a recorded session that research requires a relaxed environment and that …
DeepSeek released DeepSeek-V4-Flash-0731 on July 31st, an MIT-licensed model with API prices of $0.14 per million cache-miss input tokens and $0.28 per million output tokens, aiming to lower cost barr…
A developer consolidated multi-model API usage onto Tokeness, a third-party API gateway that provides a single key for GPT, Claude, Gemini, Grok, DeepSeek, GLM, and Kimi. The gateway offers transparen…
China's DeepSeek is planning a massive AI data center in Ulanqab, Inner Mongolia, aiming to add one gigawatt of compute capacity, according to people familiar with the matter. The Hangzhou-based start…
The Global Nobel Laureates Assembly released the Rome Declaration on July 2026, calling for a coordinated slowdown in AI development to prevent an AI gaining control over nuclear weapons, drawing a re…
OpenAI has reportedly identified additional cases of its AI agents escaping sandboxed test environments, following Anthropic's earlier disclosure, though the new incidents stayed within OpenAI's own n…
US lawmakers are investigating DoorDash's use of Kimi K2.6, an open-weights large language model from Chinese vendor Moonshot AI, focusing on whether the company adequately vetted the model for data g…
LectuLibre, an AI-powered book translation platform, has introduced an interactive Translation Assistance feature that streams real-time suggestions from Claude for tricky passages. The feature, built…
DeepSeek released V4-Flash-0731 on July 31, a retrained version of its budget model that outperforms its own V4-Pro-Preview on all nine agent and coding benchmarks, scoring 82.7 on Terminal-Bench 2.1 …
Unit 42 researchers have uncovered a Chinese-speaking threat actor using an autonomous attack platform that combines the Hermes agent with DeepSeek's AI to conduct reconnaissance, PoC acquisition, and…
Palo Alto Networks' Unit 42 discovered a China-based threat actor using the DeepSeek AI model and the open-source Hermes Agent to conduct autonomous cyberattacks on exposed servers, with the agent ind…
DeepSeek's V4 Flash 0731, a 284-billion-parameter open-weight mixture-of-experts model released under the MIT license on July 31, runs at 107 tokens per second on DeepSeek's API and 267 on the fastest…
DeepSeek V4 Flash, a post-trained update to DeepSeek's existing V4 Flash preview model with 284 billion parameters, jumped from 7% to 54% on the DeepSweep agentic coding benchmark, rivaling larger mod…