DeepSeek Upgrades R1 AI Model: Improved Reasoning, Less Hallucination, and a Distilled Version

DeepSeek updates its R1 reasoning AI model, releases it on Hugging Face | TechCrunch

Chinese startup DeepSeek has released an updated version of its R1 reasoning AI model on the developer platform Hugging Face after announcing it in a WeChat message Wednesday morning. The updated R1, which is under a permissive MIT license, meaning it can be used commercially, is a "minor" upgrade, according to DeepSeek's WeChat announcement. The Hugging Face repository doesn't contain a description of the model -- only configuration files and weights, the internal components of a model that guide its behavior. Weighing in at 685 billion parameters in size, the updated R1 is quite hefty. ("Parameters" is synonymous with "weights.") Without modification, the model likely can't run on consumer-grade hardware. DeepSeek rose to prominence earlier this year following the release of R1, which gave models from OpenAI a run for their money. The startup has raised the ire of some regulators stateside, who argue that DeepSeek's technology poses a national security risk.

DeepSeek's distilled new R1 AI model can run on a single GPU | TechCrunch

DeepSeek's updated R1 reasoning AI model might be getting the bulk of the AI community's attention this week. But the Chinese AI lab also released a smaller, "distilled" version of its new R1, DeepSeek-R1-0528-Qwen3-8B, that DeepSeek claims beats comparably-sized models on certain benchmarks. The smaller updated R1, which was built using the Qwen3-8B model Alibaba launched in May as a foundation, performs better than Google's Gemini 2.5 Flash on AIME 2025, a collection of challenging math questions. DeepSeek-R1-0528-Qwen3-8B also nearly matches Microsoft's recently released Phi 4 reasoning plus model on another math skills test, HMMT. So-called distilled models like DeepSeek-R1-0528-Qwen3-8B are generally less capable than their full-sized counterparts. On the plus side, they're far less computationally demanding. According to the cloud platform NodeShift, Qwen3-8B requires a GPU with 40GB-80GB of RAM to run (e.g., an Nvidia H100). The full-sized new R1 needs around a dozen 80GB GPUs. DeepSeek trained DeepSeek-R1-0528-Qwen3-8B by taking text generated by the updated R1 and using it to fine-tune Qwen3-8B. In a dedicated webpage for the model on the AI dev platform Hugging Face, DeepSeek describes DeepSeek-R1-0528-Qwen3-8B as "for both academic research on reasoning models and industrial development focused on small-scale models." DeepSeek-R1-0528-Qwen3-8B is available under a permissive MIT license, meaning it can be used commercially without restriction. Several hosts, including LM Studio, already offer the model through an API.

DeepSeek says a new R1 update is closing the gap with OpenAI o3 and Gemini 2.5 Pro.

The Chinese AI model that shook up the industry as a more cost-efficient alternative to the ones from OpenAI, Google, and Meta, now has a new update dubbed DeepSeek-R1-0528. DeepSeek says its latest model has a reduced "hallucination" rate, and that it "has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training...overall performance is now approaching that of leading models, such as O3 and Gemini 2.5 Pro."

DeepSeek Says Upgraded Model Reasons Better, Hallucinates Less

The Chinese startup DeepSeek said Thursday that its upgraded artificial-intelligence model can perform mathematics, programming, and general logic better than the previous version, while hallucinating less. The upgrade to its R1 model -- which stunned the AI world in January by rivaling the systems of much-larger US developers despite being built at what the Chinese startup said was a fraction of the cost -- features a greater depth of reasoning, DeepSeek said in a post on the AI model platform Hugging Face.

China's DeepSeek releases an update to its R1 reasoning model

SHANGHAI, May 29 - Chinese artificial intelligence startup DeepSeek released an update to its R1 reasoning model in the early hours of Thursday, stepping up competition with U.S. rivals such as OpenAI. DeepSeek launched R1-0528 on developer platform Hugging Face, but has yet to make an official public announcement. It did not publish a description of the model or comparisons. But the LiveCodeBench leaderboard, a benchmark developed by researchers from UC Berkeley, MIT, and Cornell, ranked DeepSeek's updated R1 reasoning model just slightly behind OpenAI's o4 mini and o3 reasoning models on code generation and ahead of xAI's Grok 3 mini and Alibaba's Qwen 3. Bloomberg earlier reported the update on Wedneday. It said that a DeepSeek representative had told a WeChat group that it had completed what it described as a "minor trial upgrade" and that users could start testing it. DeepSeek earlier this year upended beliefs that U.S. export controls were holding back China's AI advancements after the startup released AI models that were on a par or better than industry-leading models in the United States at a fraction of the cost. The launch of R1 in January sent tech shares outside China plummetting in January and challenged the view that scaling AI requires vast computing power and investment. Since R1's release, Chinese tech giants like Alibaba (9988.HK), opens new tab and Tencent (0700.HK), opens new tab have released models claiming to surpass DeepSeek's. Google's (GOOGL.O), opens new tab Gemini has introduced discounted tiers of access while OpenAI cut prices and released an o3 Mini model that relies on less computing power. The company is still widely expected to release R2, a successor to R1. Reuters reported in March, citing sources, that R2's release was initially planned for May. DeepSeek also released an upgrade to its V3 large language model in March. Reporting by Brenda Goh and Eduardo Baptista; Editing by Michael Perry Our Standards: The Thomson Reuters Trust Principles., opens new tab Suggested Topics:Artificial Intelligence

China's DeepSeek quietly releases upgraded R1 AI model, ramping up competition with OpenAI

Deepseek's logo on Jan. 29, 2025.Andrey Rudakov | Bloomberg | Getty Images Chinese startup DeepSeek, which caused shockwaves across markets this year, quietly released an upgraded version of its artificial intelligence reasoning model. The company did not make an official announcement, but the upgrade of DeepSeek R1 was released on AI model repository Hugging Face. DeepSeek rose to prominence this year after its free, open-source R1 reasoning model outperformed offerings from rivals including Meta and OpenAI. The low-cost and short time of development shocked global markets, sparking concerns that U.S. tech giants were overspending on infrastructure and wiping billions of dollars of value of major U.S. tech stocks like AI stalwart Nvidia. These companies have since broadly recovered. Just as was the case with DeepSeek R1's debut, the upgraded model was also released with little fanfare. It is a reasoning model, which means the AI can execute more complicated tasks through a step-by-step logical thought process. The upgraded DeepSeek R1 model is just behind OpenAI's o4-mini and o3 reasoning models on LiveCodeBench, a site that benchmarks models against different metrics. DeepSeek has become the poster child of how Chinese artificial intelligence is still developing despite U.S. attempts to restrict the country's access to chips and other technology. This month, Chinese technology giants Baidu and Tencent revealed how they were making their AI models more efficient to deal with U.S. semiconductor export curbs.

DeepSeek R1-0528 arrives in powerful open source challenge to OpenAI o3 and Google Gemini 2.5 Pro

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More After rocking the global AI and business community early this year with the January 20 initial release of its hit open source reasoning AI model R1, the Chinese startup DeepSeek -- a spinoff of formerly only locally well-known Hong Kong quantitative analysis firm High-Flyer Capital Management -- has released DeepSeek-R1-0528, a significant update that brings DeepSeek's free and open model near parity in reasoning capabilities with proprietary paid models such as OpenAI's o3 and Google Gemini 2.5 Pro This update is designed to deliver stronger performance on complex reasoning tasks in math, science, business and programming, along with enhanced features for developers and researchers. Like its predecessor, DeepSeek-R1-0528 is available under the permissive and open MIT License, supporting commercial use and allowing developers to customize the model to their needs. Open-source model weights are available via the AI code sharing community Hugging Face, and detailed documentation is provided for those deploying locally or integrating via the DeepSeek API. Existing users of the DeepSeek API will automatically have their model inferences updated to R1-0528 at no additional cost. The current cost for DeepSeek's API is For those looking to run the model locally, DeepSeek has published detailed instructions on its GitHub repository. The company also encourages the community to provide feedback and questions through their service email. Individual users can try it for free through DeepSeek's website here, though you'll need to provide a phone number or Google Account access to sign in. Enhanced reasoning and benchmark performance At the core of the update are significant improvements in the model's ability to handle challenging reasoning tasks. DeepSeek explains in its new model card on HuggingFace that these enhancements stem from leveraging increased computational resources and applying algorithmic optimizations in post-training. This approach has resulted in notable improvements across various benchmarks. In the AIME 2025 test, for instance, DeepSeek-R1-0528's accuracy jumped from 70% to 87.5%, indicating deeper reasoning processes that now average 23,000 tokens per question compared to 12,000 in the previous version. Coding performance also saw a boost, with accuracy on the LiveCodeBench dataset rising from 63.5% to 73.3%. On the demanding "Humanity's Last Exam," performance more than doubled, reaching 17.7% from 8.5%. These advances put DeepSeek-R1-0528 closer to the performance of established models like OpenAI's o3 and Gemini 2.5 Pro, according to internal evaluations -- both of those models either have rate limits and/or require paid subscriptions to access. UX upgrades and new features Beyond performance improvements, DeepSeek-R1-0528 introduces several new features aimed at enhancing the user experience. The update adds support for JSON output and function calling, features that should make it easier for developers to integrate the model's capabilities into their applications and workflows. Front-end capabilities have also been refined, and DeepSeek says these changes will create a smoother, more efficient interaction for users. Additionally, the model's hallucination rate has been reduced, contributing to more reliable and consistent output. One notable update is the introduction of system prompts. Unlike the previous version, which required a special token at the start of the output to activate "thinking" mode, this update removes that need, streamlining deployment for developers. Smaller variants for those with more limited compute budgets Alongside this release, DeepSeek has distilled its chain-of-thought reasoning into a smaller variant, DeepSeek-R1-0528-Qwen3-8B, which should help those enterprise decision-makers and developers who don't have the hardware necessary to run the full This distilled version reportedly achieves state-of-the-art performance among open-source models on tasks such as AIME 2024, outperforming Qwen3-8B by 10% and matching Qwen3-235B-thinking. According to Modal, running an 8-billion-parameter large language model (LLM) in half-precision (FP16) requires approximately 16 GB of GPU memory, equating to about 2 GB per billion parameters. Therefore, a single high-end GPU with at least 16 GB of VRAM, such as the NVIDIA RTX 3090 or 4090, is sufficient to run an 8B LLM in FP16 precision. For further quantized models, GPUs with 8-12 GB of VRAM, like the RTX 3060, can be used. DeepSeek believes this distilled model will prove useful for academic research and industrial applications requiring smaller-scale models. Initial AI developer and influencer reactions The update has already drawn attention and praise from developers and enthusiasts on social media. Haider aka "@slow_developer" shared on X that DeepSeek-R1-0528 "is just incredible at coding," describing how it generated clean code and working tests for a word scoring system challenge, both of which ran perfectly on the first try. According to him, only o3 had previously managed to match that performance. Meanwhile, Lisan al Gaib posted that "DeepSeek is aiming for the king: o3 and Gemini 2.5 Pro," reflecting the consensus that the new update brings DeepSeek's model closer to these top performers. Another AI news and rumor influencer, Chubby, commented that "DeepSeek was cooking!" and highlighted how the new version is nearly on par with o3 and Gemini 2.5 Pro. Chubby even speculated that the last R1 update might indicate that DeepSeek is preparing to release its long-awaited and presumed "R2" frontier model soon, as well. Looking Ahead The release of DeepSeek-R1-0528 underscores DeepSeek's commitment to delivering high-performing, open-source models that prioritize reasoning and usability. By combining measurable benchmark gains with practical features and a permissive open-source license, DeepSeek-R1-0528 is positioned as a valuable tool for developers, researchers, and enthusiasts looking to harness the latest in language model capabilities. Let me know if you'd like to add any more quotes, adjust the tone further, or highlight additional elements!

New DeepSeek-R1 Is as Good as OpenAI o3 and Gemini 2.5 Pro | AIM

The new DeepSeek-R1-0528 has "significantly improved its depth of reasoning and inference capabilities." Chinese AI model maker DeepSeek announced a new update to its R1 reasoning model on Wednesday. The updated model, DeepSeek-R1-0528, is available on Hugging Face. "In the latest update, DeepSeek R1 has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimisation mechanisms during post-training," said DeepSeek. The company also shared the model's benchmark results, which showed that it achieved performance parity with OpenAI's o3 and Google's Gemini 2.5 Pro models on multiple evaluations. In the AIME 2025 test, DeepSeek-R1-0528 scored 87.5%, close to OpenAI-o3 (88.9%) and outperformed Gemini 2.5 Pro's (83.0%). Besides, the model achieved scores on par with leading AI models on other coding, mathematics, and reasoning evaluations, as seen on Artificial Analysis. It scored 77% on LiveCodeBench (coding benchmark), matching Gemini 2.5 Pro (77%) and nearly OpenAI's o3 (78%) in coding ability. On the reasoning and general knowledge benchmark MMLU-Pro, DeepSeek-R1 achieved 85%, comparable to Gemini 2.5 Pro (84%) and OpenAI's o3 (85%). Source: Artificial Analysis Several users have already downloaded and deployed the model locally, as per their social media posts. Ivan Fioravanti, CTO of CoreView, said on X that he could run the DeepSeek-R1-0528-4bit at around 21 tokens per second on an Apple M3 Ultra chip-based device. The DeepSeek-R1 reasoning model, released last year, created quite a storm across the AI ecosystem. During its launch, the model surpassed several competing ones in benchmarks. DeepSeek prioritises using efficient techniques in the model's architecture to improve performance rather than relying on high computing power. One of DeepSeek's previous models, V3, used 2048 NVIDIA H800 GPUs to achieve performance better than most open-source models. Andrej Karpathy, former OpenAI researcher, said the DeepSeek V3's level of capability is 'supposed to require clusters of closer to 16,000 GPUs'. This caused numerous entities to doubt the demand for AI-related hardware, resulting in a market cap loss of over $500 billion for NVIDIA in just one day. Numerous startups and products use the open-source DeepSeek model for deployment, and its capabilities are extensively recognised across various sectors in China. Recently, it was reported to be used for research and development for the country's 'most advanced warplanes'. Besides, German automotive leader BMW revealed plans to incorporate DeepSeek into its vehicles in China. Last month, The New York Times revealed that courtroom officials are utilising DeepSeek to draft legal documents in minutes. Additionally, doctors and agencies are employing the model to locate missing persons. The report further noted that numerous companies are "encouraging" employees to adopt DeepSeek for design and customer service tasks.

DeepSeek releases upgraded AI model, claims parity with ChatGPT, Gemini

The R1-0528 model features enhanced reasoning capabilities, with DeepSeek asserting it's closing the gap with OpenAI's and Google's latest models. DeepSeek, a China-based artificial intelligence company, has announced an upgrade to its AI chatbot, saying it can now offer enhanced overall logic, mathematics and programming with a reduced hallucination rate. According to DeepSeek, the upgraded model -- DeepSeek-R1-0528 -- has "significantly improved its depth of reasoning and inference capabilities." The startup said the model's overall performance is now "approaching that of leading models, such as O3 and Gemini 2.5 Pro." DeepSeek's debut of its R1 chatbot in January sent shockwaves through the AI industry and further established China as an AI force. The company's first AI model had a training cost of $6 million and similar performance to leading AI models trained on significantly larger sums of capital. According to data from Business of Apps, DeepSeek has been downloaded 75 million times since its launch and had 38 million monthly active users (MAU) as of April. In a recent antitrust lawsuit, Google estimated that Gemini reached 350 million active users in March, while OpenAI's ChatGPT claimed 600 million active users in the same month. Related: China's DeepSeek launches new open-source AI after R1 took on OpenAI The United States government is planning to restrict the sale of advanced chip design software to China. According to a Bloomberg report, the move seeks to limit China's ability to advance its domestic semiconductor manufacturing capabilities. Semiconductors are critical for a wide range of technologies, including AI, where they serve as the hardware backbone for training and running complex models. New China AI models, such as Tencent's T1 and Alibaba's Qwen3, have also emerged in the first few months of 2025, spurring the AI race along.

A new DeepSeek R1 AI model just hit the Hugging Face platform

DeepSeek, a Chinese startup, released an updated version of its R1 reasoning AI model on Hugging Face Wednesday, following an announcement on WeChat. The original R1 model gained prominence earlier this year and rivaled models from OpenAI. The updated R1 model is a "minor" upgrade, according to DeepSeek's WeChat announcement. The model is available under a permissive MIT license, allowing for commercial use. The Hugging Face repository for the updated R1 model, identified as DeepSeek-R1-0528, consists of configuration files and weights, but lacks a description of the model itself. The updated R1 model contains 685 billion parameters. Due to its size, the model likely requires specialized hardware beyond consumer-grade capabilities. DeepSeek's technology has also garnered attention from regulators in the United States, with some arguing that it poses a national security risk.

DeepSeek Unveils Update to R1 Model as AI Race Heats Up

DeepSeek said it has upgraded the R1 artificial-intelligence model that helped propel the Chinese startup to global prominence earlier this year. DeepSeek completed what it described as a "minor trial upgrade" and told users they can start testing it, according to a company representative's post in an official WeChat group on Wednesday. The Hangzhou-based startup stunned the global tech industry in January when it unveiled R1, an AI model that outperformed Western players on several standardized metrics, purportedly at a cost of just several million dollars. That triggered a rout in global tech stocks as investors questioned whether leading firms would still need to spend significant amounts to build AI services. The debut of R1 also set off a race to launch additional AI models in China. Founder Liang Wenfeng became a symbol of the country's ability to compete with the best of Silicon Valley. © 2025 Bloomberg LP

DeepSeek unveils update to R1 model

DeepSeek has announced a minor upgrade to its R1 AI model, which previously stunned the tech world by outperforming Western competitors at a fraction of the cost. This news arrives shortly before Nvidia's latest financial report, whose shares were impacted by the initial R1 release. The upgrade follows founder Liang Wenfeng's rise to prominence and recognition from President Xi Jinping.DeepSeek said it has upgraded the R1 artificial-intelligence model that helped propel the Chinese startup to global prominence earlier this year. DeepSeek completed what it described as a "minor trial upgrade" and told users they can start testing it, according to a company representative's post in an official WeChat group on Wednesday. The company didn't provide details about the upgrade and didn't respond to an email seeking further comment. The Hangzhou-based startup stunned the global tech industry in January when it unveiled the original R1, an AI model that outperformed Western players on several standardised metrics, purportedly at a cost of just several million dollars. That triggered a rout in global tech stocks as investors questioned whether leading firms would still need to spend significant amounts to build AI services. The debut of R1 turned founder Liang Wenfeng into a tech celebrity and a symbol of the country's ability to compete with the best of Silicon Valley. It also set off a race to launch additional AI models in China. In February, President Xi Jinping invited Liang to a high-profile gathering with some of the country's most prominent entrepreneurs. The young founder was seated among the likes of Alibaba Group Holding Ltd. co-founder Jack Ma and Tencent Holdings Ltd.'s Pony Ma. DeepSeek's upgrade was announced just hours before the latest financial report from Nvidia Corp., the leading maker of AI chips whose shares were pummeled in the January rout.

China's DeepSeek releases an update to its R1 reasoning model

DeepSeek earlier this year upended beliefs that U.S. export controls were holding back China's AI advancements after the startup released AI models that were on a par or better than industry-leading models in the United States at a fraction of the cost.Chinese artificial intelligence startup DeepSeek released an update to its R1 reasoning model in the early hours of Thursday, stepping up competition with U.S. rivals such as OpenAI. DeepSeek launched R1-0528 on developer platform Hugging Face, but has yet to make an official public announcement. It did not publish a description of the model or comparisons. But the LiveCodeBench leaderboard, a benchmark developed by researchers from UC Berkeley, MIT, and Cornell, ranked DeepSeek's updated R1 reasoning model just slightly behind OpenAI's o4 mini and o3 reasoning models on code generation and ahead of xAI's Grok 3 mini and Alibaba's Qwen 3. Bloomberg earlier reported the update on Wedneday. It said that a DeepSeek representative had told a WeChat group that it had completed what it described as a "minor trial upgrade" and that users could start testing it. DeepSeek earlier this year upended beliefs that U.S. export controls were holding back China's AI advancements after the startup released AI models that were on a par or better than industry-leading models in the United States at a fraction of the cost. The launch of R1 in January sent tech shares outside China plummetting in January and challenged the view that scaling AI requires vast computing power and investment. Since R1's release, Chinese tech giants like Alibaba and Tencent have released models claiming to surpass DeepSeek's. Google's Gemini has introduced discounted tiers of access while OpenAI cut prices and released an o3 Mini model that relies on less computing power. The company is still widely expected to release R2, a successor to R1. reported in March, citing sources, that R2's release was initially planned for May. DeepSeek also released an upgrade to its V3 large language model in March.

DeepSeek Upgrades AI Reasoning Model to Rival OpenAI and Google | PYMNTS.com

The model also hallucinates less, has enhanced support for function calling and offers a better experience in vibe coding, where developers use natural language prompts in an AI chatbot to write code, per the post. Read also: DeepSeek Debuts Upgrade to AI Model That Improves Reasoning and Coding As an open-source model, R1-0528 is free to use. It offers an MIT license, which lets users download, run and modify the model. Cloud providers Amazon Web Services (AWS) and Microsoft Azure offer DeepSeek's R1 model to their clients as part of their respective AI platforms. But they strip out any connection to Chinese servers, so data stays in the client's chosen servers, AWS has told PYMNTS. Companies with their own developers can also use DeepSeek's R1-0528 model and customize it for their own use cases. As long as they don't use DeepSeek's API, the data stays in their designated servers. DeepSeek and Meta offer some of the most popular and powerful open-source AI models as a counter to proprietary models offered by OpenAI, Google, Microsoft, Anthropic and others. While open-source models are free, they are not always cheaper to use depending on how much customization is needed. A company that doesn't have the staff to customize would still need to hire an outside firm to do so. Also, running the model incurs costs per token in the cloud, unless users can host the model on their own servers.

China's DeepSeek Upgrades R1 Reasoning Model with New Features

DeepSeek's R1 Update Reinforces China's AI Push Amid Export Controls Chinese AI startup DeepSeek has quietly slipped in an update to its R1 reasoning model. This update occurred as a competitive move against US AI developers like OpenAI. The new update, R1-0528, silently appeared on the messaging platform Hugging Face without an official announcement or a hazy description of its functionalities. However, the benchmark results show that DeepSeek's model has reached a close second to OpenAI's finest, being the o4 mini and o3 reasoning models, in code generation performances.