DeepSeek releases open-source AI models that rival GPT-5 and Gemini at fraction of the cost

DeepSeek Releases New Reasoning Models to Take On ChatGPT and Gemini

Imad is a senior reporter covering Google and internet culture. Hailing from Texas, Imad started his journalism career in 2013 and has amassed bylines with The New York Times, The Washington Post, ESPN, Tom's Guide and Wired, among others. DeepSeek, the China-based AI company, has released two new reasoning-capable AI models, V3.2 and V3.2-Speciale, with the latter outperforming Gemini 3.0 Pro and GPT-5 High in benchmarks, the company said in news release on Monday. At the moment, V3.2 is available on the app and web, whereas V3.2-Speciale can only be accessed via the API. DeepSeek V3.2 is a follow-up to V3.2-Exp, which stands for experimental and was released in September. The Hangzhou-based AI company says V3.2 surpasses the performance of OpenAI's GPT-5 in reasoning benchmarks and has reasoning performance "on par" with Gemini 3.0 Pro. DeepSeek didn't release testing against GPT-5 Pro. That could be because ChatGPT is blocked in China. What's more, DeepSeek said it's committed to staying open source, meaning any company can load DeepSeek's models for free. But considering V3.2 has 685 billion parameters, it can only run on a giant server stack with millions of dollars in hardware. Don't miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source. DeepSeek didn't immediately respond to a request for comment. DeepSeek's release of 3.2 comes as the company continues to put American AI companies on notice. The company first gained worldwide attention earlier this year when it released a reasoning-capable model online, for free. It immediately turned the narrative about reasoning models on its head, showing that it's possible to run smarter, more capable AI models at a fraction of the cost. This spooked Wall Street, with some wondering if OpenAI, Google and Anthropic weren't innovating fast enough. Since then, ChatGPT, Gemini and Claude have all released reasoning-level models for free, with higher-level "thinking" models being available for paid subscribers. DeepSeek-V3.2-Speciale won gold in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI), the company said. (Disclosure: Ziff Davis, CNET's parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

Is DeepSeek's new model the latest blow to proprietary AI?

Chinese AI firm DeepSeek has made yet another splash with the release of V3.2, the latest iteration in its V3 model series. Launched Monday, the model, which builds on an experimental V3.2 version announced in October, comes in two versions: "Thinking," and a more powerful "Speciale." DeepSeek said V3.2 pushes the capabilities of open-source AI even further. Like other DeepSeek models, it's a fraction of the cost of proprietary models, and the underlying weights can be accessed via Hugging Face. Also: I tested DeepSeek's R1 and V3 coding skills - and we're not all doomed (yet) DeepSeek first made headlines in January with the release of R1, an open-source reasoning AI model that outperformed OpenAI's o1 on several crucial benchmarks. Considering V3.2's performance also rivals powerful proprietary models, could this shake up the AI industry once more? Rumors first began circulating in September that DeepSeek was planning to launch its own, more cost-effective agent to compete with the likes of OpenAI and Google. Now, it seems that the competitor has finally arrived. V3.2 is the latest iteration of V3, a model DeepSeek released nearly a year ago that also helped inform R1. According to company data published Monday, V3.2 Speciale outperforms industry-leading proprietary models like OpenAI's GPT-5 High, Anthropic's Claude 4.5 Sonnet, and Google's Gemini 3.0 Pro on some reasoning benchmarks (for what it's worth, Kimi K2, a free and open-source model from Moonshot, also claimed to rival GPT-5 and Sonnet 4.5 in performance). In terms of cost, accessing Gemini 3 in the API costs up to $4.00 per 1 million tokens, while V3.2 Speciale is $0.028 per 1 milion tokens. The new model also achieved gold-level performance in the International Math Olympiad (IMO) and the International Olympiad in Informatics, according to the company. "DeepSeek-V3.2 emerges as a highly cost-efficient alternative in agent scenarios, significantly narrowing the performance gap between open and frontier proprietary models while incurring substantially lower costs," the company wrote in a research paper. While these claims are still debated, the sentiment continues DeepSeek's pattern of reducing costs with each model release, which threatens to logically undermine the exorbitant investments proprietary labs like OpenAI pour into their models. DeepSeek said it built V3.2 in an effort to help the open-source AI community catch up with some of the technical achievements that have recently been made by companies building closed-source models. According to the company's paper, the agentic and reasoning capabilities demonstrated by leading proprietary models have "accelerated at a significantly steeper rate" than those of their open-source counterparts. Also: Mistral's latest open-source release bets on smaller models over large ones - here's why As the engineer Charles Kettering once put it, "A problem well-stated is a problem half-solved." In that spirit, DeepSeek began the development of its new model by attempting to diagnose the reasons behind open-source models' lagging performance, ultimately breaking it down into three factors. First, open-source models have tended to rely on what's known to AI researchers as "vanilla attention" -- a slow and compute-hungry mechanism for reading inputs and generating outputs, which makes them struggle with longer sequences of tokens. They also have a more computationally limited post-training phase, hindering their ability to complete more complex tasks. Unlike proprietary models, they struggle with following long instructions and generalizing across tasks, making them inefficient agents. In response, the company introduced DeepSeek Sparse Attention (DSA), a mechanism that mitigates "critical computation complexity without sacrificing long-context performance," according to the research paper. Also: What is sparsity? DeepSeek AI's secret, revealed by Apple researchers With traditional vanilla attention, a model essentially generates its outputs by comparing each individual token from a query with every single token in its training data -- a painstakingly power-hungry process. By analogy, imagine you had to dig through an enormous pile of books scattered on a lawn to find a particular sentence. You could do it, but it would take a lot of time and careful scrutiny of a huge number of pages. The DSA approach tries to work smarter, not harder. It's deployed in two phases: an initial "lightning indexer" search, which performs a high-level scan of the tokens in its training data to identify the small subset that are likely to be most relevant to a particular query. It then drills into that subset with its full computational power to find what it's looking for. Rather than starting with a giant pile of books, you're now able to walk into a neatly organized library, walk to the relevant section, and perform a much less stressful and lengthy search to find the passage you've been seeking. The company then aimed to solve the post-training issue by building "specialist" models to test and refine V3.2's abilities across writing, general question-answering, mathematics, programming, logical reasoning, agentic tasks, agentic coding, and agentic search. They're like tutors charged with the task of turning the model from a generalist into a multi-specialist. DeepSeek V3.2, according to the research paper, "effectively bridges the gap between computational efficiency and advanced reasoning capabilities" and "[unlocks] new possibilities for robust and generalizable AI agents" through open-source AI. Also: Stop saying AI hallucinates - it doesn't. And the mischaracterization is dangerous There are a few caveats, however. For one thing, the new model's "world knowledge" -- the breadth of practical understanding about the real world that can be inferred from a corpus of training data -- is much more limited compared to leading proprietary models. It also requires more tokens to generate outputs that match the quality of those from frontier proprietary models, and it struggles with more complex tasks. DeepSeek says it plans to continue bridging the divide between its own open-source models and its proprietary counterparts by scaling up compute during pretraining and refining its "post-training recipe." Even with these limitations, though, the fact that a company -- and one based in China, no less -- has built an open-source model that can compete with the reasoning capabilities of some of the most advanced proprietary models currently on the market is a huge deal. It reiterates growing evidence that the "performance gap" between open-source and close-sourced models isn't a fixed and unresolvable fact, but a technical discrepancy that can be bridged through creative approaches to pretraining, attention, and posttraining. Even more importantly, the fact that its underlying weights are almost free for developers to access and build upon could undermine the basic sales pitch that's thus far been deployed by the industry's leading developers of closed-source models: That it's worth paying to access these tools, since they're the best on the market. If open-source models eclipse proprietary models, it won't make sense for most people to continue paying for the latter.

DeepSeek just gave away an AI model that rivals GPT-5 - and it could change everything

Their release challenges U.S. tech dominance and reignites global AI competition Chinese startup DeepSeek has casually lobbed two new enormous AI models into the already smoldering international arms race: DeepSeek-V3.2, a model built for everyday reasoning, and DeepSeek-V3.2-Speciale, a high-octane variant that allegedly outperformed top American models in elite math and coding competitions. Not only that, but they released them under an open-source license. What makes this move notable isn't just the models and their capabilities, but also how they were released. American firms like OpenAI and Google rely on powerful, and often expensive, models relying on private APIs and red-team testing for the most cutting-edge models. DeepSeek has weaponized openness. DeepSeek-V3.2 reportedly matches or beats GPT-5 and Gemini 3 Pro on long-form reasoning, tool use, and dense problem solving, including competitions like the International Mathematical Olympiad and the ICPC World Finals. The "Speciale" version scored 99.2% on the Harvard-MIT Math Tournament, 73% on software bug fixing, and posted gold-medal results on multiple international benchmarks even without internet access or external tools. The trick behind this performance is a clever architectural hack called DeepSeek Sparse Attention, or DSA. Traditional transformer models become computationally bloated as context length increases, and they have to consider every word in a document relative to every other word. DSA reduces costs by focusing only on the most relevant parts of the input, essentially skimming rather than reading every word. That alone slashes costs for long documents by up to 70%, making the model relatively cheap. This matters to real people because cost is everything in AI deployment. Most frontier models today are trapped behind paywalls and throttled access. But DeepSeek's latest models and their 128,000-token context windows are free to download and modify. A solo developer or student team can tinker with systems that just a few months ago would've required a lab and a huge cloud budget. DeepSeek's "thinking in tool-use" breakthrough is especially notable. Most AI agents struggle to juggle multiple tools because each action resets their internal reasoning. DeepSeek fixed that by preserving memory across tools. The company trained the model using over 85,000 complex synthetic instructions to make it work with tools like real web browsers and coding environments. That's a level of real-world task preparation that most current chatbots simply aren't built for. It's one thing to summarize a recipe. It's another thing to plan a multi-day vacation under a strict budget with interdependent lodging and food constraints and to do it while testing code snippets and checking exchange rates. The license setup might be even more disruptive. By using the MIT open-source license, DeepSeek has made it legally possible for anyone to copy, remix, or commercialize its models. That flies directly in the face of the current trend of protecting model weights as intellectual property, citing safety, misuse risk, and corporate secrecy. Openness doesn't mean transparency, however. That's why German regulators have tried to block DeepSeek over data transfer concerns. Italy banned the app earlier this year, and U.S. lawmakers want it off government devices entirely. DeepSeek is a Chinese company, and the geopolitical context looms large. But suppose DeepSeek's models really do deliver frontier performance at a fraction of the cost, and you don't mind the geopolitical baggage. What exactly are the American firms offering that's worth the markup? For now, DeepSeek's Speciale variant is only available via a temporary API. But by mid-December, it will be merged into the broader V3.2 release and accessible to everyone. If the last few years were defined by ChatGPT's friendly intro to AI, this release feels like a stark reminder: the gloves are off, and the global AI race isn't just about features anymore, it's about access, cost, and control.

DeepSeek just dropped two insanely powerful AI models that rival GPT-5 and they're totally free

Chinese artificial intelligence startup DeepSeek released two powerful new AI models on Sunday that the company claims match or exceed the capabilities of OpenAI's GPT-5 and Google's Gemini-3.0-Pro -- a development that could reshape the competitive landscape between American tech giants and their Chinese challengers. The Hangzhou-based company launched DeepSeek-V3.2, designed as an everyday reasoning assistant, alongside DeepSeek-V3.2-Speciale, a high-powered variant that achieved gold-medal performance in four elite international competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, the ICPC World Finals, and the China Mathematical Olympiad. The release carries profound implications for American technology leadership. DeepSeek has once again demonstrated that it can produce frontier AI systems despite U.S. export controls that restrict China's access to advanced Nvidia chips -- and it has done so while making its models freely available under an open-source MIT license. "People thought DeepSeek gave a one-time breakthrough but we came back much bigger," wrote Chen Fang, who identified himself as a contributor to the project, on X (formerly Twitter). The release drew swift reactions online, with one user declaring: "Rest in peace, ChatGPT." How DeepSeek's sparse attention breakthrough slashes computing costs At the heart of the new release lies DeepSeek Sparse Attention, or DSA -- a novel architectural innovation that dramatically reduces the computational burden of running AI models on long documents and complex tasks. Traditional AI attention mechanisms, the core technology allowing language models to understand context, scale poorly as input length increases. Processing a document twice as long typically requires four times the computation. DeepSeek's approach breaks this constraint using what the company calls a "lightning indexer" that identifies only the most relevant portions of context for each query, ignoring the rest. According to DeepSeek's technical report, DSA reduces inference costs by roughly half compared to previous models when processing long sequences. The architecture "substantially reduces computational complexity while preserving model performance," the report states. Processing 128,000 tokens -- roughly equivalent to a 300-page book -- now costs approximately $0.70 per million tokens for decoding, compared to $2.40 for the previous V3.1-Terminus model. That represents a 70% reduction in inference costs. The 685-billion-parameter models support context windows of 128,000 tokens, making them suitable for analyzing lengthy documents, codebases, and research papers. DeepSeek's technical report notes that independent evaluations on long-context benchmarks show V3.2 performing on par with or better than its predecessor "despite incorporating a sparse attention mechanism." The benchmark results that put DeepSeek in the same league as GPT-5 DeepSeek's claims of parity with America's leading AI systems rest on extensive testing across mathematics, coding, and reasoning tasks -- and the numbers are striking. On AIME 2025, a prestigious American mathematics competition, DeepSeek-V3.2-Speciale achieved a 96.0% pass rate, compared to 94.6% for GPT-5-High and 95.0% for Gemini-3.0-Pro. On the Harvard-MIT Mathematics Tournament, the Speciale variant scored 99.2%, surpassing Gemini's 97.5%. The standard V3.2 model, optimized for everyday use, scored 93.1% on AIME and 92.5% on HMMT -- marginally below frontier models but achieved with substantially fewer computational resources. Most striking are the competition results. DeepSeek-V3.2-Speciale scored 35 out of 42 points on the 2025 International Mathematical Olympiad, earning gold-medal status. At the International Olympiad in Informatics, it scored 492 out of 600 points -- also gold, ranking 10th overall. The model solved 10 of 12 problems at the ICPC World Finals, placing second. These results came without internet access or tools during testing. DeepSeek's report states that "testing strictly adheres to the contest's time and attempt limits." On coding benchmarks, DeepSeek-V3.2 resolved 73.1% of real-world software bugs on SWE-Verified, competitive with GPT-5-High at 74.9%. On Terminal Bench 2.0, measuring complex coding workflows, DeepSeek scored 46.4% -- well above GPT-5-High's 35.2%. The company acknowledges limitations. "Token efficiency remains a challenge," the technical report states, noting that DeepSeek "typically requires longer generation trajectories" to match Gemini-3.0-Pro's output quality. Why teaching AI to think while using tools changes everything Beyond raw reasoning, DeepSeek-V3.2 introduces "thinking in tool-use" -- the ability to reason through problems while simultaneously executing code, searching the web, and manipulating files. Previous AI models faced a frustrating limitation: each time they called an external tool, they lost their train of thought and had to restart reasoning from scratch. DeepSeek's architecture preserves the reasoning trace across multiple tool calls, enabling fluid multi-step problem solving. To train this capability, the company built a massive synthetic data pipeline generating over 1,800 distinct task environments and 85,000 complex instructions. These included challenges like multi-day trip planning with budget constraints, software bug fixes across eight programming languages, and web-based research requiring dozens of searches. The technical report describes one example: planning a three-day trip from Hangzhou with constraints on hotel prices, restaurant ratings, and attraction costs that vary based on accommodation choices. Such tasks are "hard to solve but easy to verify," making them ideal for training AI agents. DeepSeek employed real-world tools during training -- actual web search APIs, coding environments, and Jupyter notebooks -- while generating synthetic prompts to ensure diversity. The result is a model that generalizes to unseen tools and environments, a critical capability for real-world deployment. DeepSeek's open-source gambit could upend the AI industry's business model Unlike OpenAI and Anthropic, which guard their most powerful models as proprietary assets, DeepSeek has released both V3.2 and V3.2-Speciale under the MIT license -- one of the most permissive open-source frameworks available. Any developer, researcher, or company can download, modify, and deploy the 685-billion-parameter models without restriction. Full model weights, training code, and documentation are available on Hugging Face, the leading platform for AI model sharing. The strategic implications are significant. By making frontier-capable models freely available, DeepSeek undermines competitors charging premium API prices. The Hugging Face model card notes that DeepSeek has provided Python scripts and test cases "demonstrating how to encode messages in OpenAI-compatible format" -- making migration from competing services straightforward. For enterprise customers, the value proposition is compelling: frontier performance at dramatically lower cost, with deployment flexibility. But data residency concerns and regulatory uncertainty may limit adoption in sensitive applications -- particularly given DeepSeek's Chinese origins. Regulatory walls are rising against DeepSeek in Europe and America DeepSeek's global expansion faces mounting resistance. In June, Berlin's data protection commissioner Meike Kamp declared that DeepSeek's transfer of German user data to China is "unlawful" under EU rules, asking Apple and Google to consider blocking the app. The German authority expressed concern that "Chinese authorities have extensive access rights to personal data within the sphere of influence of Chinese companies." Italy ordered DeepSeek to block its app in February. U.S. lawmakers have moved to ban the service from government devices, citing national security concerns. Questions also persist about U.S. export controls designed to limit China's AI capabilities. In August, DeepSeek hinted that China would soon have "next generation" domestically built chips to support its models. The company indicated its systems work with Chinese-made chips from Huawei and Cambricon without additional setup. DeepSeek's original V3 model was reportedly trained on roughly 2,000 older Nvidia H800 chips -- hardware since restricted for China export. The company has not disclosed what powered V3.2 training, but its continued advancement suggests export controls alone cannot halt Chinese AI progress. What DeepSeek's release means for the future of AI competition The release arrives at a pivotal moment. After years of massive investment, some analysts question whether an AI bubble is forming. DeepSeek's ability to match American frontier models at a fraction of the cost challenges assumptions that AI leadership requires enormous capital expenditure. The company's technical report reveals that post-training investment now exceeds 10% of pre-training costs -- a substantial allocation credited for reasoning improvements. But DeepSeek acknowledges gaps: "The breadth of world knowledge in DeepSeek-V3.2 still lags behind leading proprietary models," the report states. The company plans to address this by scaling pre-training compute. DeepSeek-V3.2-Speciale remains available through a temporary API until December 15, when its capabilities will merge into the standard release. The Speciale variant is designed exclusively for deep reasoning and does not support tool calling -- a limitation the standard model addresses. For now, the AI race between the United States and China has entered a new phase. DeepSeek's release demonstrates that open-source models can achieve frontier performance, that efficiency innovations can slash costs dramatically, and that the most powerful AI systems may soon be freely available to anyone with an internet connection. As one commenter on X observed: "Deepseek just casually breaking those historic benchmarks set by Gemini is bonkers." The question is no longer whether Chinese AI can compete with Silicon Valley. It's whether American companies can maintain their lead when their Chinese rival gives comparable technology away for free.

DeepSeek Releases New Reasoning Models to Match GPT-5, Rival Gemini 3 Pro | AIM

Both models and the accompanying technical report have been released as open source on Hugging Face. Chinese AI lab DeepSeek has launched two new reasoning-first AI models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, expanding its suite of systems for agents, tool-use and complex inference. Both the models and the accompanying technical report have been released as open source on Hugging Face. The company announced on X that V3.2 is the official successor to V3.2-Exp and is now available across its app, web interface and API. The Speciale variant is offered only through a temporary API endpoint until December 15, 2025. DeepSeek said V3.2 aims to balance inference efficiency with long-context performance, calling it "your daily driver at GPT-5 level performance." The V3.2-Speciale model, positioned for high-end reasoning tasks, "rivals Gemini-3.0-Pro," the company said. According to DeepSeek, Speciale delivers gold-level (expert human proficiency) results across competitive benchmarks such as the IMO, CMO and ICPC World Finals. The models introduce an expansion of DeepSeek's agent-training approach, supported by a new synthetic dataset spanning more than 1,800 environments and 85,000 complex instructions. The company stated that V3.2 is its first model to integrate thinking directly into tool use, allowing structured reasoning to operate both within and alongside external tools. Alongside the release, DeepSeek updated its API, noting that V3.2 maintains the same usage pattern as its predecessor. The Speciale model is priced the same as V3.2 but does not support tool calls. The company also highlighted a new capability in V3.2 described as "Thinking in Tool-Use," with additional details provided in its developer documentation. The company recently also released a new open-weight model, DeepSeekMath-V2. The model, as per the AI lab, demonstrates strong theorem-proving capabilities in mathematics and achieved gold-level scores on the International Mathematics Olympiad (IMO) 2025.

How DeepSeek 3.2 Uses Specialists & Improved Memory to Outthink Gemini 3.0 Pro

What if we told you that an open source AI just outperformed one of the most advanced proprietary models on the market? Yes, you read that right -- DeepSeek 3.2 has officially outclassed Gemini 3.0 Pro, a feat many thought impossible just a few years ago. For decades, open source AI has been seen as the underdog, often dismissed as a step behind its corporate-funded counterparts. But with new innovations like Sparse Attention and domain-specific training, DeepSeek 3.2 has shattered those expectations, proving that open systems can deliver not just parity but superiority. This isn't just an incremental improvement; it's a paradigm shift that could redefine the future of artificial intelligence. In this deep dive, Universe of AI explore how DeepSeek 3.2 has tackled the core challenges that have historically held open source AI back -- computational inefficiency, reasoning gaps, and agent behavior limitations -- and turned them into strengths. You'll discover how its innovative features, from enhanced memory retention to domain-specific expertise, are allowing it to excel in everything from debugging code to solving Olympiad-level math problems. But what does this mean for the broader AI landscape? Could this be the tipping point where open source models finally rival, or even surpass, the proprietary giants? Let's unpack the innovations, implications, and real-world applications that make DeepSeek V3.2 a fantastic option. The development of DeepSeek 3.2 directly tackles three persistent challenges that have historically limited the competitiveness of open source AI models: computational inefficiency, weak reasoning capabilities, and limitations in agent behavior. These obstacles have long hindered open models from excelling in tasks requiring advanced reasoning, long-context processing, and multi-step planning. DeepSeek 3.2 introduces a suite of innovative features designed to overcome these challenges, positioning it as a formidable competitor to leading proprietary AI systems. These advancements not only enhance the model's performance but also expand its practical applications. Check out more relevant guides from our extensive collection on DeepSeek 3 that you might find useful. DeepSeek 3.2 has demonstrated exceptional performance across a range of benchmarks, often rivaling or surpassing proprietary models. Its achievements underscore the potential of open source AI to deliver competitive results in both theoretical and practical domains. The release of DeepSeek 3.2 signifies a fantastic moment for open source AI, proving that accessible models can achieve performance levels once thought exclusive to proprietary systems. This achievement not only sets a new standard for innovation but also fosters collaboration and accessibility within the AI community.

DeepSeek's New AI Model Achieves Gold-Level Results, Rivals GPT-5 and Gemini 3 Pro

The DeepSeek V3.2-Speciale model delivers gold-level results in IMO, CMO, ICPC, and IOI 2025. China's frontier AI lab, DeepSeek has released its new reasoning AI models that rival the performance of OpenAI's GPT-5 High and Google's Gemini 3 Pro. There are two AI models including DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The DeepSeek-V3.2 model is the successor to the V3.2-Experimental model and it powers the DeepSeek app. The more powerful, DeepSeek-V3.2-Speciale is quite performant and it's only available via the API. This special model achieves gold-level results in IMO, CMO, ICPC World Finals, and IOI 2025, something that OpenAI and Google have achieved with their specialized models. A variant of Gemini 2.5 Deep Think achieved the Gold medal at IMO 2025. While it's far more capable, it also consumes much more token and there is no tool support for this model yet. In terms of agentic capabilities too, the new models are much better. Now, coming to benchmarks, the DeepSeek-V3.2-Speciale model achieves 96.0 in AIME 2025 whereas GPT-5 High gets 94.6 and Gemini 3 Pro stands at 95.0. In Humanity's Last Exam (HLE), the new special model gets 30.6 whereas Gemini 3 Pro achieves 37.7. Now, in SWE Verified, DeepSeek's new model achieves 73.1, a bit lower than Gemini 3 Pro (76.2). All in all, the new model by DeepSeek is quite frontier-class and it highlights that China is a strong contender in the AI race. Along with Alibaba's Qwen 3 AI models, now DeepSeek is delivering strong performance across various benchmarks.

DeepSeek 'Integrates Thinking' Into New AI Models | PYMNTS.com

By completing this form, you agree to receive marketing communications from PYMNTS and to the sharing of your information with our sponsor, if applicable, in accordance with our Privacy Policy and Terms and Conditions. "DeepSeek-V3.2 is our first model to integrate thinking directly into tool-use, and also supports tool-use in both thinking and non-thinking modes," the Chinese startup wrote in a post on X Monday (Dec. 1). The launch was flagged in a report by Bloomberg news, which noted that these releases are new versions of a model released weeks ago, but with new capabilities designed to help combine reasoning and execute certain actions autonomously. Per that report, DeepSeek said the new services match the performance of OpenAI's flagship GPT-5 throughout multiple reasoning benchmarks, which indicates China's open-source systems are still competitive with Silicon Valley's models on at least some metrics. DeepSeek also says the V3.2 version combines an ability to mimic something akin to human reasoning with the ability to use search engines, calculators and code executors. DeepSeek shook the AI sector at the start of this year with the debut of an AI model that it said could perform at the level of many U.S. flagship models but had been developed at just a fraction of the cost. "Artificial intelligence has reached a critical inflection point. The industry stands at a crossroads where escalating costs, environmental concerns, and innovation appear intertwined, threatening to stifle accessibility and adoption," Gokul Naidu, a consultant for SAP, told PYMNTS soon after the initial launch. "Enter DeepSeek-R1, the model that's turning heads in Silicon Valley and beyond for proving that high performance and affordability aren't mutually exclusive." "DeepSeek challenges the narrative that innovation must come at an unsustainable cost," Naidu added. "For businesses, this means AI could soon be accessible to small and medium enterprises, not just tech giants with deep pockets." In other AI news, PYMNTS wrote about the progression of enterprise AI adoption three years after the debut of OpenAI's ChatGPT. Research from PYMNTS Intelligence shows that companies have begun treating AI as a critical part of their workforce, and not just an experiment. Sixty large U.S. companies are reorganizing roles and responsibilities under new chief AI officers. The research showed that 34% of CFOs pointed to increased output as their chief reason for adopting AI, followed by staying competitive at 24% and improved decision-making via better data insights, cited by 19% of respondents. "The data also revealed sharp industry divides," PYMNTS wrote. "48% of goods sector firms are using AI to boost output and efficiency. In contrast, 30% of service firms aimed to improve decision-making and customer experience with AI. Additionally, 42% of technology firms stated that their main goal is to maintain their competitive edge."

DeepSeek launches pair of AI models to challenge OpenAI, Google

China-based DeepSeek (DEEPSEEK) has launched a pair of new artificial intelligence models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, which are open-sourced and topped or matched the results of OpenAI's (OPENAI) GPT-5 and Google's (GOOG)(GOOGL) Gemini 3 Pro in some benchmarks. DeepSeek-V3.2 is now DeepSeek-V3.2 and V3.2-Speciale matched or outperformed GPT-5 and Gemini 3 Pro in some benchmarks, notably in mathematical and informatics Olympiad tasks. DeepSeek introduced DSA for efficient long-sequence attention, a scalable RL framework, and a large-scale agentic task synthesis pipeline. DeepSeek cites inefficient vanilla attention mechanisms, lack of post-training compute investment, and lagging generalization and instruction-following abilities.

DeepSeek 3.2 Challenges GPT-5 While Slashing AI Spend : IMO Gold to 128k Context

What if innovative AI didn't have to come with a sky-high price tag? Imagine an open source model that not only rivals proprietary giants like GPT-5 but also delivers gold medal-level performance in global competitions, all while being 30 times more affordable. Enter DeepSeek 3.2, a new leap in artificial intelligence that's shaking up the status quo. With its ability to tackle complex reasoning tasks and dominate benchmarks in mathematics, programming, and more, DeepSeek is proving that you don't need to break the bank to access world-class AI. But can it truly compete with the closed-source titans of the industry? And what does this mean for the future of open source innovation? In this overview AI Grid explains how DeepSeek 3.2 is redefining the AI landscape by balancing affordability with performance. From its innovative Sparse Attention technology to its unmatched cost-efficiency, this model is designed to empower developers, researchers, and organizations across the globe. Yet, it's not without its trade-offs, limitations in efficiency and specialized features reveal the challenges of open source development. Whether you're curious about its gold medal achievements, its practical applications, or its potential to provide widespread access to AI access, this deep dive will uncover what makes DeepSeek 3.2 a fantastic option, and where it still has room to grow. After all, the real question isn't just how it compares to GPT-5, but how it's reshaping what we expect from AI itself. DeepSeek 3.2 and its enhanced counterpart, DeepSeek 3.2 Special, are designed to meet a wide range of user needs, offering flexibility and accessibility: As open source solutions, these models provide widespread access to access to advanced AI technologies. By eliminating the high costs typically associated with proprietary systems, DeepSeek aligns with a growing industry trend toward inclusivity and accessibility. This approach enables users across various sectors to use high-performance AI without financial barriers, fostering innovation and collaboration. DeepSeek's models have demonstrated exceptional capabilities across a variety of benchmarks, showcasing their strengths in reasoning, programming, and mathematical problem-solving: These achievements underscore the models' ability to handle complex reasoning tasks with precision. However, certain limitations remain. For instance, the models are less efficient and lack some specialized features offered by closed-source competitors like Gemini 3 Pro and Opus 4.5. These gaps highlight areas for future development, particularly in improving real-world usability and operational efficiency. Check out more relevant guides from our extensive collection on DeepSeek that you might find useful. DeepSeek 3.2 incorporates several advanced technologies that significantly improve its performance and operational efficiency: These innovations enable the models to excel in multi-step reasoning tasks and integrate seamlessly with external tools. This versatility makes them suitable for a wide range of applications, from legal document analysis to large-scale data synthesis. By focusing on efficiency and adaptability, DeepSeek 3.2 positions itself as a practical solution for both individual developers and large organizations. One of the standout features of DeepSeek 3.2 is its ability to process extensive context inputs, up to 128,000 tokens, without significantly increasing operational costs. This capability is particularly valuable for tasks requiring deep contextual understanding, such as analyzing lengthy legal documents, conducting academic research, or synthesizing large datasets. Additionally, the models are up to 30 times more cost-effective than GPT-5, offering a compelling alternative for budget-conscious users. However, the Special model's advanced reasoning capabilities come with higher token usage, which can offset its cost advantages for resource-intensive applications. This trade-off emphasizes the importance of selecting the appropriate model based on specific use cases and operational requirements. For users prioritizing affordability, the standard DeepSeek 3.2 model offers a balanced solution, while the Special model caters to those needing advanced functionality. Despite their impressive capabilities, DeepSeek's models face several challenges that highlight the trade-offs inherent in open source AI development: These limitations reflect the ongoing balancing act between accessibility and performance in open source systems. Addressing these challenges will be crucial for DeepSeek to expand its user base and compete more effectively with proprietary alternatives. The release of DeepSeek 3.2 and 3.2 Special represents a pivotal moment in the AI industry. By delivering competitive performance at a fraction of the cost, these models challenge the dominance of established players like OpenAI and Anthropic. This development is particularly significant for emerging markets and non-Western regions, where affordability is a critical factor in technology adoption. Additionally, the success of DeepSeek's models highlights the growing contributions of non-Western developers to the global AI landscape. This shift toward a more diverse and inclusive industry reflects the increasing widespread access of AI technologies, paving the way for broader innovation and collaboration. As open source AI continues to evolve, models like DeepSeek 3.2 are poised to play an increasingly influential role in shaping the future of artificial intelligence. By prioritizing affordability, accessibility, and performance, these models exemplify the potential of open source systems to drive meaningful progress in the field.

DeepSeek 3.2 AI Outperforms GPT-5 & Gemini 3 Thanks to a New Bold Training Method

What does it take to outshine giants in the fiercely competitive world of artificial intelligence? For years, proprietary systems like GPT-5 and Gemini Pro have dominated the landscape, setting benchmarks others could only chase. Yet, against all odds, DeepSeek has done it again. With the release of DeepSeek 3.2 and its enhanced sibling, DeepSeek 3.2 Special, the company has redefined what open-weight AI systems can achieve. From outperforming industry titans in logic and mathematics to earning accolades in global competitions, these models are proving that innovation doesn't always come with a closed label. But how did they pull it off, and what challenges still stand in their way? Below Prompt Engineering unpacks the new advancements that make DeepSeek 3.2 Special a standout in the AI world. You'll discover how innovative attention mechanisms and domain-specific training have pushed the boundaries of reasoning and problem-solving, and why these models are being hailed as a fantastic option for scientific research and advanced analytics. Yet, the story isn't without its complexities, challenges like token efficiency and ecosystem limitations reveal the uphill battle open-weight systems still face. As we delve into the mechanics and implications of DeepSeek's latest triumph, one question lingers: can open-weight AI truly rival its proprietary counterparts in shaping the future of intelligence? DeepSeek 3.2 Highlights DeepSeek 3.2 and its enhanced counterpart, DeepSeek 3.2 Special, represent a significant leap in AI capabilities. Both models excel in reasoning and problem-solving, but DeepSeek 3.2 Special is specifically optimized for tasks requiring advanced logic and mathematical precision. This optimization has earned it accolades, including gold medals in prestigious global competitions like the International Math Olympiad. These achievements underscore the models' ability to handle complex, structured problems with exceptional accuracy. The enhanced version, DeepSeek 3.2 Special, is particularly adept at tasks requiring intricate reasoning and mathematical rigor, making it a preferred choice for applications in scientific research, engineering, and advanced analytics. These distinctions highlight the growing versatility and specialization of open-weight AI systems in addressing diverse challenges. Key Innovations Driving DeepSeek's Success DeepSeek's advancements are underpinned by several new innovations that address critical challenges in AI development. These innovations enable the models to achieve high performance while maintaining efficiency and scalability: * DeepSeek Sparse Attention (DSA): This mechanism dynamically selects relevant tokens, allowing the models to process long contexts more efficiently. By reducing computational costs without compromising performance, DSA is a crucial factor in scaling AI systems for real-world applications. * Reinforcement Learning (RL): DeepSeek allocates a significant portion of its compute resources to reinforcement learning during post-training, 10% for DeepSeek 3.2 and 20% for DeepSeek 3.2 Special. This approach enhances the models' reasoning and decision-making capabilities by training them in synthetic environments to solve complex problems. * Domain-Specific Training: Employing a "divide and conquer" strategy, DeepSeek uses specialized teacher models for distinct domains. These teacher models distill their expertise into a unified general model, making sure consistent performance across a wide range of tasks. These innovations collectively enable DeepSeek's models to excel in areas such as logic, mathematics, and coding, setting them apart from their competitors. DeepSeek 3.2 Special Performance and Training Gains Explained Below are more guides on DeepSeek AI models from our extensive range of articles. Performance Benchmarks and Competitive Edge DeepSeek 3.2 and 3.2 Special have established new benchmarks in reasoning, logic, and problem-solving. They consistently outperform GPT-5 and Gemini Pro in key areas, particularly in tasks involving mathematics, logic, and coding. Notably, DeepSeek 3.2 Special achieved GPT-5-level reasoning in mathematics even before GPT-5's public release, demonstrating its advanced capabilities. In addition to their technical achievements, these models offer a lower cost per token, making them more accessible for specific applications. However, token efficiency remains a challenge, as the models require longer token generation paths to achieve comparable quality to proprietary systems. This trade-off highlights the ongoing need for optimization in open-weight AI systems. Challenges and Areas for Improvement Despite their impressive capabilities, DeepSeek's models face several limitations. Open-weight systems inherently have fewer training resources compared to proprietary models, resulting in a narrower breadth of general knowledge. While their performance in specialized tasks is commendable, they still lag behind leading closed systems in terms of general-purpose functionality. Another challenge lies in token efficiency. The reliance on longer token generation paths can hinder real-time applications, limiting the models' scalability in commercial settings. Addressing these limitations will be critical for DeepSeek to expand its impact and compete more effectively with proprietary systems. Innovative Tool Usage and Ecosystem Challenges DeepSeek has introduced a novel approach to tool usage during the reasoning process. By discarding historical thinking traces when new user inputs are introduced, the models optimize their performance for specific scenarios. This approach enhances their adaptability and precision in dynamic environments. However, the ecosystem supporting open-weight models remains underdeveloped compared to proprietary systems like Gemini and OpenAI. This disparity limits the scalability and integration of DeepSeek's models in commercial applications. Building a robust ecosystem will be essential for maximizing the potential of open-weight AI systems. Hardware Integration and Future Opportunities The release of DeepSeek 3.2 and 3.2 Special also highlights the potential for synergy between AI software and hardware. With the emergence of advanced chips like Huawei Ascend, open-weight models could benefit from hardware optimizations that enhance performance and reduce costs. This integration may help level the playing field against proprietary systems, opening new opportunities for innovation and competition in the AI industry. DeepSeek's focus on hardware compatibility underscores its commitment to advancing the capabilities of open-weight AI systems. By using innovative hardware, the company aims to overcome existing limitations and unlock new possibilities for its models. Future Directions for DeepSeek DeepSeek has outlined ambitious plans to address its current limitations and further refine its models. Key areas of focus include improving token efficiency, expanding pre-training compute resources, and enhancing foundational models. Additionally, the company plans to develop more advanced post-training strategies to strengthen the reasoning and problem-solving capabilities of its models. These efforts reflect DeepSeek's dedication to driving innovation in the AI space. By addressing existing challenges and building on its strengths, the company aims to solidify its position as a leader in open-weight AI development.

New Deepseek 3.2 AI Open Model Outthinks ChatGPT 5 in Tough Reasoning Tests

What if the next leap in artificial intelligence wasn't locked behind corporate walls, but instead, freely available to everyone? That's the bold promise of Deepseek 3.2, the latest evolution in open source AI. With its jaw-dropping gold medal wins at both the International Math Olympiad (IMO) and the International Olympiad in Informatics (IOI), this isn't just another incremental update, it's a seismic shift. Deepseek 3.2 doesn't just compete with industry titans like GPT-5 and Gemini 3.0 Pro; in some areas, it outright surpasses them. From its new sparse attention mechanisms to its ability to tackle complex, multi-step reasoning tasks, this release has redefined what open source AI can achieve. But what makes Deepseek 3.2 truly remarkable isn't just its performance, it's its accessibility and scalability. Whether you're a researcher, developer, or simply curious about the future of AI, this open source powerhouse offers something for everyone. In the video below Matthew Berman explains how features like reinforcement learning and agentic task synthesis push the boundaries of reasoning and adaptability, while innovations like linear-scaling sparse attention make it more efficient than ever. How does it manage to rival, and in some cases outshine, its closed-source competitors? And what does this mean for the future of provide widespread access tod AI? Let's unpack the breakthroughs that are reshaping the landscape of artificial intelligence and putting the power back in the hands of its users. Deepseek 3.2 has achieved several milestones that underscore its leadership in the AI domain. These accomplishments highlight its ability to excel in both theoretical and practical applications: Deepseek 3.2 integrates a suite of advanced technologies that enhance its performance while maintaining scalability and efficiency. These features are designed to address the growing demands of modern AI applications. The introduction of the Deepseek Sparse Attention (DSA) mechanism is a cornerstone of Deepseek 3.2's efficiency. Unlike traditional attention systems that scale quadratically, DSA employs a more linear scaling approach. This allows the model to process longer context windows without sacrificing performance. By reducing computational complexity, DSA not only enhances processing speed but also lowers operational costs, making the model more accessible for diverse use cases. Deepseek 3.2 allocates over 10% of its compute resources to post-training reinforcement learning. This deliberate investment improves the model's ability to generalize across tasks and follow instructions with greater precision. The reinforcement learning framework equips Deepseek 3.2 to adapt to a wide array of challenges, making sure consistent performance across diverse scenarios. The agentic task synthesis pipeline is another new feature of Deepseek 3.2. By using 1,800 environments and generating 85,000 complex prompts, this pipeline enables the model to train on a vast and varied dataset. This approach significantly enhances its reasoning and problem-solving capabilities, particularly in tasks that require agentic behavior and effective tool use. Explore further guides and articles from our vast library that you may find relevant to your interests in Deepseek AI. Deepseek 3.2 is engineered to deliver high performance while optimizing resource utilization. Its technical specifications reflect a balance between scalability and precision: Deepseek 3.2 excels in tasks that require advanced reasoning, decision-making, and tool utilization. Its ability to narrow the performance gap between open source and closed-source models in tool-use benchmarks is a testament to its robust design. This capability makes it an invaluable resource for applications that demand precision, adaptability, and the integration of external tools. Designed with scalability and cost-efficiency as core principles, Deepseek 3.2 is accessible to a wide audience, from academic researchers to industry professionals. Its open weights and open source licensing ensure that it can be freely used and adapted for various purposes. By combining state-of-the-art reasoning capabilities with an accessible framework, Deepseek 3.2 plays a crucial role in providing widespread access to AI technology and fostering innovation across the global AI community. Deepseek 3.2 stands as a benchmark in the evolution of open source AI, offering unparalleled advancements in reasoning, efficiency, and scalability. Its innovative features, such as sparse attention mechanisms, reinforcement learning, and agentic task synthesis, position it as a leader in the field. By prioritizing accessibility and collaboration, Deepseek 3.2 not only bridges the gap between open source and proprietary models but also paves the way for future breakthroughs in artificial intelligence.

DeepSeek introduces new AI models that rival GPT-5 and Gemini 3.0: Here's what it offers

The launch marks DeepSeek's closest move yet toward AGI, after earlier versions shook global tech stocks. After Google and OpenAI introduced Gemini 3.0 and GPT-5, China's AI startup DeepSeek has introduced two new large language models, DeepSeek-V3.2 and DeepSeek-V3.2 Speciale, in order to keep pace with the rapid advancements in the AI race. Taking to the blog post, DeepSeek described them as "reasoning-first" systems built with a focus on efficiency and advanced tool-use capabilities. As per the company, the standard V3.2 model builds on earlier experimental versions, while the Speciale edition is made to deliver strong reasoning performance. The company said V3.2 is its first model to integrate reasoning directly into tool interactions and is capable of generating large-scale agent-training data across more than 1,800 environments and over 85,000 complex instructions. Both models use the company's DeepSeek Sparse Attention (DSA) mechanism, which is specifically designed to lower computational costs while preserving performance on longer inputs. The newly introduced V3.2 model, as claimed by the company, matches the performance that OpenAI's GPT-5 offers. On the other hand, the company says the Speciale variant is positioned as the direct rival to Google's Gemini 3.0-Pro. The Speciale model has also reportedly achieved gold-medal-level results in the 2025 International Mathematical Olympiad and the International Olympiad in Informatics, two benchmarks often used to evaluate high-end reasoning systems. DeepSeek previously got massive attention earlier this year worldwide after the reports of its strong model performance caused major US tech stocks to drop. The company also stood out by offering AI tools at much lower costs than its rivals. It remains to be seen how the LLMs perform.