DeepSeek R1
DeepSeek R1 is an open-weight reasoning model trained primarily via reinforcement learning. It reproduced much of the performance of OpenAI’s o1 at a fraction of the inference price.
Released
January 20, 2025
Type
reasoning
License
open-weight
Context
128,000 tokens
Pricing
Input
$0.55 / 1M tokens
Output
$2.19 / 1M tokens
Benchmarks
| AIME 2024 | 79.8% |
| GPQA Diamond | 71.5% |
Capabilities
Architecture
Parameters: 685B MoE (37B active)
Links
DeepSeek R1 in the news
Fortune · Sep 13, 2026
Faced with less compute and fewer tokens, Chinese AI labs are tightening the gap with the U.S. by just being more efficient
US government agencies have accused six Chinese artificial intelligence companies of illegally acquiring capabilities worth billions of dollars by purchasing bulk subscriptions to American AI models and training on their outputs. The FBI, NSA, and CISA stated on Tuesday that this method allowed companies like DeepSeek to significantly reduce their training costs, with DeepSeek allegedly understating its training expenses to $5.6 million. China's foreign affairs ministry has refuted these accusations, asserting that the nation's AI advancements are a product of self-reliance and deeming the allegations "groundless." These allegations offer one possible explanation for the competitive standing of Chinese AI models, which a Stanford report earlier this year indicated were nearly on par with their US counterparts, with Anthropic's top model leading DeepSeek's by only 2.7%. However, industry analysts suggest that Chinese AI laboratories have also developed a distinct advantage by mastering techniques to extract greater value from limited computational resources and data. This efficiency is particularly evident in their optimization of the "attention" mechanism, a core component of large language models introduced by Google researchers in 2017. The attention mechanism allows AI models to weigh the importance of different parts of input text to understand context, but its computational cost increases significantly with longer context windows. Brendan Burke, a semiconductor and supply chain analyst at Futurum Group, explained to Fortune that Chinese labs have innovated algorithms that reduce the complexity of these attention calculations, making their models more efficient. This focus on algorithmic efficiency allows Chinese AI developers to achieve high performance with less compute power and fewer tokens, a strategy that contrasts with the often resource-intensive development approaches seen in the US. The implication is that while accusations of intellectual property theft are being investigated, the more sustainable and perhaps more significant factor in China's rapid AI progress is its ingenuity in optimizing existing technologies and maximizing the utility of available resources. This efficiency-driven approach enables Chinese AI companies to compete effectively on a global scale, even when facing constraints in compute power and data availability compared to their US rivals. The ongoing debate highlights the complex landscape of AI development, encompassing both competitive pressures and the pursuit of technological advancement through diverse strategies.
The Hacker News · Sep 11, 2026
Anthropic Says Seven China-Based AI Labs Ran Industrial-Scale Claude Distillation Attacks
Anthropic announced on Thursday that it has identified and disrupted industrial-scale illicit distillation attacks targeting its Claude AI model. These sophisticated attacks were traced back to seven distinct artificial intelligence laboratories based in China. The identified entities include prominent organizations such as Alibaba, Moonshot, DeepSeek, Z.ai (also known as Zhipu), and MiniMax. The attacks involved knowledge distillation, a legitimate machine learning technique where a larger, more capable AI model acts as a 'teacher' to train a smaller model. However, in this context, the labs were allegedly using Anthropic's powerful Claude models to illicitly train their own AI systems without authorization, bypassing Anthropic's terms of service and potentially infringing on intellectual property rights. Anthropic's security team detected these activities by monitoring for unusual patterns of access and data exfiltration that were indicative of large-scale, systematic attempts to replicate Claude's capabilities. The company stated that these distillation efforts were "industrial-scale," suggesting a significant investment of resources and computational power by the implicated labs. Such attacks aim to extract proprietary knowledge and model parameters from a teacher model to create a student model that mimics its performance, often at a fraction of the training cost and time. By engaging in this unauthorized distillation, the Chinese labs sought to gain a competitive advantage by rapidly developing advanced AI models based on Anthropic's research and development. In response to the discovery, Anthropic took immediate action to mitigate the threat. The company stated that it has taken steps to block the identified actors and prevent further unauthorized access and misuse of its Claude models. This disruption aims to protect Anthropic's intellectual property and maintain the integrity of its AI systems. The company emphasized that knowledge distillation is a valid technique when conducted ethically and with proper licensing, but the actions of these seven labs constituted a violation of their acceptable use policies. The incident highlights the ongoing challenges in securing advanced AI models against sophisticated intellectual property theft and unauthorized replication attempts, particularly in a competitive global AI landscape. Anthropic has not disclosed the specific versions of Claude that were targeted in these attacks, nor the exact methods used by the labs to extract the necessary data for distillation. However, the scale and systematic nature of the attacks suggest a coordinated effort. The company's proactive stance in identifying and reporting these activities underscores the increasing importance of AI security and the need for robust defenses against emerging threats. The involvement of multiple well-known Chinese AI companies also points to the intense competition and rapid development occurring within China's AI sector, where the pursuit of cutting-edge capabilities may sometimes lead to questionable practices. Anthropic's disclosure serves as a warning to other AI developers and a signal of the evolving security landscape in artificial intelligence.
Bloomberg Markets · Sep 11, 2026
DeepSeek's New Model Rattles Chipmakers and AI Rivals
DeepSeek has launched Coder 2.5, an updated iteration of its large language model specifically engineered for code generation tasks. This new model represents a substantial advancement over its predecessor, Coder 2.0, and aims to provide developers with enhanced capabilities for writing, debugging, and optimizing software. The release positions DeepSeek as a significant contender in the competitive landscape of AI-powered development tools, challenging established players and offering a compelling alternative for businesses and individual programmers. Coder 2.5 is built upon DeepSeek's proprietary architecture, which has been refined to improve its understanding of programming languages and complex coding logic. The model's training data has been expanded and curated to include a wider array of programming languages, frameworks, and coding paradigms, enabling it to generate more accurate and contextually relevant code. This comprehensive training approach is crucial for addressing the diverse needs of modern software development, which often involves multiple languages and intricate project structures. The model's performance benchmarks, as detailed in DeepSeek's technical documentation, indicate a notable increase in its ability to handle intricate coding challenges and produce functional code snippets with greater efficiency. The improvements in Coder 2.5 are expected to have a tangible impact on developer productivity. By automating repetitive coding tasks, assisting in debugging complex issues, and suggesting more efficient algorithms, the model can significantly reduce the time and effort required for software development cycles. This enhanced efficiency can translate into faster product releases, reduced development costs, and the ability for development teams to focus on more innovative aspects of their projects. Furthermore, Coder 2.5's improved understanding of natural language prompts allows developers to describe their desired functionality in plain English, with the AI translating these requirements into executable code, thereby lowering the barrier to entry for less experienced programmers. DeepSeek's strategic release of Coder 2.5 underscores the rapid evolution of AI in the software development sector. The company's commitment to continuous improvement and its focus on specialized AI models for distinct tasks, such as code generation, highlight a growing trend in the AI industry. As AI models become more sophisticated and accessible, they are increasingly integrated into the core workflows of technology companies, driving innovation and reshaping the future of software engineering. The competitive pressure from models like Coder 2.5 encourages further research and development across the board, benefiting the entire developer community.
Bloomberg Markets · Sep 11, 2026
DeepSeek Hands Memory Stock Investors Fresh Reason for Caution
Fund managers re-engaging with South Korean memory manufacturers' stocks encountered renewed apprehension following the release of DeepSeek's latest artificial intelligence model. This new AI model has introduced significant doubts regarding the current strength of demand for memory chips, a critical component in numerous electronic devices and AI hardware. The implications of this development are particularly pertinent for South Korea, a global leader in memory chip production, with companies like Samsung Electronics and SK Hynix being major players in the market. The AI model's assessment suggests a potential oversupply or a slowdown in the anticipated growth of demand, which could directly affect the revenue and profitability of these semiconductor giants. DeepSeek, a research organization focused on AI, has been developing advanced AI models, and its latest output is being closely scrutinized by financial analysts and investors. The specific details of the AI model's findings that led to this demand concern have not been fully disclosed, but the market reaction indicates a significant level of worry among those invested in the memory sector. This caution is a departure from recent optimism that had begun to build as the global economy showed signs of recovery and the demand for AI-driven technologies continued to surge. Investors had been anticipating a rebound in the memory market, which has experienced cyclical downturns in the past. However, DeepSeek's analysis introduces a new variable, prompting a reassessment of future market conditions. The potential impact extends beyond the immediate stock prices of memory chip producers. A sustained downturn in memory chip demand could have ripple effects across the broader technology industry, affecting device manufacturers, cloud service providers, and other sectors reliant on semiconductor components. Analysts are now tasked with deciphering the precise nature of the demand concerns highlighted by DeepSeek's AI model and determining whether this represents a temporary blip or a more fundamental shift in market dynamics. The performance of South Korean memory stocks will be a key indicator in the coming weeks and months, as investors and industry observers await further data and analysis to clarify the outlook for the global memory market. The situation underscores the increasing influence of AI development not only on technological capabilities but also on financial markets and investment strategies, particularly in sectors heavily reliant on technological advancements and consumer demand.
Decrypt · Sep 10, 2026
DeepSeek's New Model Nearly Matches GPT-6 Astra on Design—at 1.4% of the Cost
OpenDesign, a platform for evaluating AI models, has released results from a comparative analysis of 13 distinct AI models tasked with performing design tasks. The benchmark study revealed that DeepSeek V4.1 Flash achieved a performance score that was only 1.5 percentage points lower than that of GPT-6 Astra, a model developed by OpenAI. This close performance parity was achieved at a substantially reduced cost, with DeepSeek V4.1 Flash being approximately 70 times cheaper to operate than GPT-6 Astra. The evaluation focused on specific design-related functionalities, aiming to quantify the practical capabilities of these advanced AI systems in creative and technical design applications. DeepSeek V4.1 Flash, developed by the AI research organization DeepSeek, is a large language model designed for efficiency and performance. Its inclusion in the OpenDesign benchmark highlights its growing competitiveness against leading models from major AI labs. GPT-6 Astra, while not yet publicly released or detailed by OpenAI, is understood to be a next-generation model building upon the capabilities of its predecessors, likely incorporating advanced reasoning and multimodal processing. The comparison by OpenDesign provides an early, albeit specific, indicator of Astra's potential performance and DeepSeek's ability to offer comparable results at a fraction of the cost. The OpenDesign platform employs a standardized set of design tasks to ensure a consistent evaluation methodology across all tested models. The specific metrics used to derive the performance scores are proprietary to OpenDesign but are designed to reflect real-world design workflows and outcomes. The significant cost difference noted between DeepSeek V4.1 Flash and GPT-6 Astra suggests potential implications for the accessibility and widespread adoption of advanced AI in design industries. Organizations seeking to integrate AI into their design processes may find cost-effective solutions like DeepSeek V4.1 Flash to be a more viable option, especially if the performance gap remains minimal for their specific use cases. This comparative analysis underscores the dynamic nature of the AI development landscape, where specialized models can emerge to challenge the dominance of larger, more resource-intensive systems. The results from OpenDesign's evaluation will likely inform future development strategies for AI companies and guide purchasing decisions for businesses looking to leverage AI for design tasks. Further details on the specific design tasks, the scoring methodology, and the performance of the other 11 AI models evaluated are expected to be released by OpenDesign in subsequent reports.