#deepseekr1

Read the full report on -

DeepSeek-R1 Jailbreak: How One AI Model Built a Bypass for Itself and Other Systems

After four months, Alibaba's new model family may surpass DeepSeek-R1, the top open-weights big language model.

Qwen 3: Faster, Deeper

Overview

Qwen3 is the latest big language model from Qwen. Qwen3-235B-A22B flagship model exceeds DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro in math, coding, and general capabilities. A tiny MoE model, Qwen3-30B-A3B, beats QwQ-32B with ten times as many active parameters, and even Qwen3-4B can compete with Qwen2.5-72B-Instruct.

We are open-weighting two MoE models: Qwen3-235B-A22B, a big model with 235 billion total parameters and 22 billion activated parameters, and Qwen3-30B-A3B, a smaller model with 30 billion total parameters and 3 billion activated parameters.

Six dense models—Qwen3-32B, Qwen3-14B, Qwen3-8B, Qwen3-4B, Qwen3-1.7B, and Qwen3-0.6B—are also open-weighted under Apache 2.0.

Hugging Face, ModelScope, and Kaggle now provide post-trained and pre-trained models like Qwen3-30B-A3B-Base. It recommends SGLang and vLLM for deployment. Ollama, LMStudio, MLX, llama.cpp, and KTransformers are recommended for local usage. These solutions make Qwen3 easy to integrate into development, production, and research workflows.

Qwen 3 allows researchers, developers, and organisations worldwide to design unique solutions using these cutting-edge models.

Try Qwen3 on the mobile app and chat.qwen.ai!

Important Features

Mixed Thinking

Qwen3 models introduce hybrid problem-solving. They offer two modes:

Thinking Mode: The model deliberates before responding. This is ideal for complex topics that require more thought.

Non-Thinking Mode: The model replies almost rapidly, making it suitable for simpler questions where depth is less important than speed.

As previously established, Qwen 3 delivers smooth and scalable performance benefits connected to computational reasoning budget. This design makes task-specific budgets easier to configure, improving inference quality and cost.

Supports several languages

Qwen 3 models accommodate 119 dialects. Due to their multilingual capabilities, these models may be used worldwide, opening up new possibilities.

Increased Agentic Capability

It optimised Qwen 3 models for coding and agentic capabilities and strengthened MCP support. The following examples show how Qwen3 thinks and acts.

In comparison to Qwen2.5

Qwen3 has a much larger pretraining dataset than Qwen2.5. Qwen2.5 was pre-trained on 18 trillion tokens, whereas Qwen3 uses 36 trillion over 119 languages and dialects. Qwen2.5-VL applied these research to enhance it. To add math and code data, Qwen2.5-Math and Qwen2.5-Coder developed synthetic data. Code samples, textbooks, and Q&As are included.

Qwen3 Pre-workout

It takes three stages to prepare for training. The model was pretrained on about 30 trillion tokens with a 4K context length in stage 1 (S1). The model learnt basic language and general knowledge at this time. In stage 2 (S2), we added STEM, coding, and reasoning challenges to the dataset. The model was pretrained with 5 trillion extra tokens. High-quality long-context data was used to extend the context to 32K tokens in the last stage. This assures the model can efficiently handle longer inputs.

Qwen 3 dense base models perform similarly to Qwen2.5 base models with more parameters due to model architectural advancements, more training data, and more efficient training methods. Qwen2.5-3B/7B/14B/32B/72B-Base and Qwen3-1.7B/4B/8B/14B/32B-Base work similarly. Qwen 3 dense base models outperform Qwen2.5 models in STEM, coding, and reasoning. For Qwen3-MoE basis models, they perform similarly to Qwen2.5 dense base models with 10% of active parameters. Thus, training and inference costs drop dramatically.

Post-training

The hybrid model, which can reason step-by-step and respond swiftly, was trained using a four-stage pipeline. This pipeline includes reasoning-based reinforcement learning (RL), thinking mode fusion, long chain-of-thought (CoT) cold start, and generic RL.

First, it improved the models using lengthy CoT data from coding, maths, logical reasoning, and STEM issues. Teaching the model fundamental thinking was the goal. The second phase increased reinforcement learning computing power using rule-based incentives to better model exploration and exploitation.

The third phase enhanced the thinking model utilising extended CoT data and regularly used instruction-tuning data to include non-thinking skills. The second stage's upgraded thinking model produced this data, ensuring smooth reasoning and rapid reaction times. The fourth step employed reinforcement learning (RL) on over 20 broad-domain tasks to increase the model's general capabilities and repair undesired behaviours. Agent capabilities, format following, and instruction following were among these duties.

Agentic uses

Qwen 3 calls tools well. To fully exploit Qwen3's agentic features, use Qwen-Agent. Qwen-Agent's inherent encapsulation of tool-calling templates and parsers simplifies development.

DeepSeek - Is the top provider of the latest AI language models & business solutions. Experience advanced artificial intelligence tech for your venture needs. DeepSeek AI was made by a Chinese AI company residing in Hangzhou, Zhejiang & is an Open AI project.

Artificial intelligence is developing unprecedentedly, with companies competing to create increasingly advanced language models. One of the newest and most intriguing players in this market is DeepSeek AI, a Chinese technology company founded by Liang Wenfeng. Its goal is to provide AI solutions that can compete with products from giants such as OpenAI, Google, and Meta.

DeepSeek AI rose to prominence with DeepSeek-R1, an AI model that offers advanced reasoning and natural language processing capabilities and is available as open source. This makes DeepSeek AI an attractive alternative to proprietary solutions from Western companies.

Contents:

History and Development of DeepSeek AI

Impact on investors and the race of giants

DeepSeek: A Catalyst for Global AI Development

Why does the DeepSeek R1 stand out from other AIs?

What is DeepSeek? – Summary

History and Development of DeepSeek AI

DeepSeek AI was founded in Hangzhou in 2023 by Liang Wenfeng, who previously co-founded the High-Flyer hedge fund. From the very beginning, DeepSeek creators had a clear mission: to create efficient and accessible AI models that can be used by the global community.

One of DeepSeek AI’s distinguishing features is its approach to openness – unlike many competitors, DeepSeek creators choose to make its models open-source. This allows developers and researchers to test, develop, and deploy them without having to pay high licensing fees.

The first breakthrough was the introduction of DeepSeek-V3, which gained recognition among AI programmers and analysts in 2024. The real breakthrough, however, was the DeepSeek-R1 model, presented in January 2025.

Models and technologies

DeepSeek AI focuses on developing advanced language models that can compete with OpenAI and Google. Their flagship product is DeepSeek-R1, which is efficient, low-cost, and open-source.

The R1 offers:

Advanced language comprehension skills – comparable to GPT-4 and Gemini 1.5, Cost optimization – fewer parameters, yet high response precision, Accessible to the open source community – enabling broad-scale adaptation and implementation, Integration with business applications – thanks to API flexibility and adaptability to various sectors.

With lower operating costs, DeepSeek AI can offer more affordable solutions for businesses and developers. This poses a serious challenge to OpenAI and Google, which have dominated the commercial language model market so far.

Impact on investors and the race of giants

The emergence of Deep Seek AI has changed the dynamics of the global AI market. Particularly significant are the declines in the value of shares of technology giants such as:

Nvidia, which has so far dominated the supply of GPUs used to train AI models, Alphabet (Google), who’s Gemini has been facing criticism for context errors, Microsoft has invested billions of dollars in OpenAI and ChatGPT's integration with Windows and Office.

DeepSeek AI shows that it is possible to create equally advanced models at a lower cost, forcing competitors to adapt their strategies and find new ways to maintain dominance.

DeepSeek: A Catalyst for Global AI Development

DeepSeek is significantly accelerating AI research worldwide, setting new standards, and lowering entry barriers for smaller companies and research institutions. By using innovative training methods that achieve comparable results using significantly fewer GPUs – DeepSeek R1 was trained using just 2,000 units instead of the traditional 16,000 used by competing models – this technology enables faster iterations, modifications, and implementation of new solutions, which translates into significant reductions in operating costs. DeepSeek - As a result, the global race of giants such as OpenAI, Microsoft, and Google is gaining new momentum, forcing industry leaders to revise their investment and technology strategies to meet growing market demands.

DeepSeek’s groundbreaking technology, described by some analysts as a kind of “Sputnik moment” for AI, not only inspires increased openness and collaboration in the research community but also democratizes access to advanced AI tools, enabling faster development of innovations on a global scale. As a result, thanks to significantly lower implementation and operating costs, DeepSeek becomes a catalyst for new projects and research that can be implemented faster and more efficiently, which significantly accelerates the pace of development of the entire artificial intelligence sector worldwide, creating new opportunities for both businesses and academia.

Controversies and challenges

Despite its successes, DeepSeek AI faces numerous challenges, especially in the context of legal regulation and censorship in China. Chinese laws on political content, censorship, and data privacy are much stricter than in Western countries.

The key challenges for DeepSeek AI are:

Balancing innovation with compliance with Chinese regulations,

Limited availability outside of China – some Western companies may be wary of using Chinese AI models, Geopolitical risk that may affect the possibility of cooperation with international partners.

Additionally, the open-source DeepSeek AI model brings security challenges – the openness of the code may lead to the technology being misused, e.g. to generate disinformation or cyberattacks.

Censorship in DeepSeek

DeepSeek implements strict censorship mechanisms that aim to limit access to politically sensitive content, as required by Chinese authorities. The DeepSeek AI model, while characterized by high efficiency and an innovative approach to reasoning, automatically blocks answers to questions about topics such as the events in Tiananmen Square or the status of Taiwan.

This mechanism, based on built-in security barriers, means that when you try to ask questions about controversial issues, DeepSeek interrupts the answer generation process or redirects the discussion to neutral topics.

Although this approach by DeepSeek creators has been criticized for limiting freedom of speech and the independence of information, it is seen as necessary due to strict state regulations. However, thanks to the transparent censorship mechanism, DeepSeek AI users can see when and why the model decides to do such things, which on the one hand increases the transparency of the AI's operation, and on the other - poses a challenge to developers who want to use the AI model's full potential in global applications.

Future and prospects

Despite these challenges, DeepSeek AI plans to continue to grow and expand its global presence. The company is focused on:

Improving AI models so they can better compete with OpenAI and Google, Democratizing access to artificial intelligence, which will allow smaller companies to benefit from AI technology, and Collaboration with international partners, which can accelerate the company's global expansion!

DeepSeek AI has the potential to become one of the innovation leaders in the AI ​​sector, provided it manages to meet regulatory and business challenges.

Why does the DeepSeek R1 stand out from other AIs?

DeepSeek R1 sets new standards in AI with its operational and cost efficiency, a huge advantage over traditional AI models. First, the DeepSeek AI model was trained using just 2,000 GPUs —far fewer than the thousands used by competing systems like ChatGPT—achieving comparable results at a fraction of the cost.

Deep Seek - This resource optimization not only reduces training costs but also enables faster iterations and improvements, which is especially important in a rapidly evolving research environment. Second, DeepSeek R1 is open-source, enabling the research and development community to freely explore, modify, and integrate the AI model into a variety of projects.

This approach democratizes access to advanced AI technologies, enabling smaller companies and research institutions to develop innovative solutions without having to invest in huge capital. Another important aspect is DeepSeek R1’s unique visible “chain of thought” mechanism, which allows DeepSeek users to see the model’s reasoning steps before providing a final answer – a feature that increases transparency and builds trust, which is especially valuable in applications that require precise results, such as mathematics or programming.

DeepSeek vs. ChatGPT

DeepSeek R1, thanks to its innovative architecture and unique technological solutions, achieves results that in many benchmarks place it on par or even higher than leading American models such as OpenAI o1 or ChatGPT-4o. According to independent tests, DeepSeek R1 achieved as much as 90% accuracy in solving complex mathematical problems, while ChatGPT-4o achieved about 83% - a difference that not only emphasizes DeepSeek R1's superiority in logical-mathematical tasks but also its ability to generate solutions on its own.

In the Codeforces programming competition, this AI model ranked in the 96.3 percentile, making it an effective tool for developers and engineers, especially in tasks requiring fast and accurate debugging and algorithm creation. Moreover, in general knowledge tests such as MMLU, DeepSeek R1 achieved a result of 90.8%, which confirms its versatility and ability to understand a wide range of topics. An additional advantage of DeepSeek R1 is its exceptionally low cost – the DeepSeek AI model was trained at an estimated cost of around $5.6 million, which is just a fraction of the outlays incurred by its American counterparts, and the costs of using the API are even 20–50 times lower than in the case of ChatGPT, making it an attractive choice for companies and institutions with limited budgets.

DeepSeek vs. Microsoft

Microsoft, a global leader in AI technology investment and integration, sees the potential of DeepSeek as a way to reduce costs while expanding AI service offerings. The company has already deployed DeepSeek R1 on the Azure AI Foundry platform, enabling enterprises to easily integrate advanced yet cost-effective AI solutions into their systems. Deep Seek - Despite the huge investment in its own models and its partnership with OpenAI, Microsoft emphasizes that DeepSeek’s innovation—especially its ability to operate at significantly lower compute costs—is inspiring for the entire sector, and CEO Satya Nadella’s comments about the model’s “super impressive” approach indicate its strategic importance for the further development of the AI ​​ecosystem.

What is DeepSeek? – Summary

DeepSeek AI is a company that has quickly become one of the most important players in the artificial intelligence market. Thanks to low costs, high efficiency, and sharing models under an open-source license, DeepSeek models can realistically compete with Western giants.