DeepSeek Unveils V3.1: A Leap Forward in AI with Domestic Chip Optimization

DeepSeek Touts Model That Outdoes Flagship in Agentic AI Step

DeepSeek unveiled an update to an older model that it says surpasses the seminal R1 on key benchmarks, keeping the Chinese startup in the game while the industry awaits its next flagship offering. The V3.1 returns answers to queries much faster and marks the startup's first step toward creating an AI agent, DeepSeek said in a WeChat post Thursday. DeepSeek first outlined the V3.1 earlier this week, but the platform only just made it to the Hugging Face portal. The version has been customized to work with next-generation Chinese-made AI chips, DeepSeek said in a separate message.

DeepSeek's new V3.1 hints at potent new Chinese chips

Point release retuned with new FP8 datatype for better compatibility with homegrown silicon Chinese AI darling DeepSeek unveiled an update to its flagship large language model that the company claims is already optimized for use with a new generation of homegrown silicon. According to DeepSeek, it trained the new V3.1 model using the UE8M0 data type, scaling the FP8 format that's already supported by the likes of Nvidia. In a WeChat comment, the org clarified that the change was made in anticipation of a new generation of silicon. "UE8M0 FP8 is designed for the next generation of domestically produced chips to be released soon," the company wrote. Lower-precision data types offer several benefits, including reduced memory consumption and higher throughput for both inference and training. However, it's worth noting DeepSeek was already using FP8, specifically the E4M3 type. As such, the switch to UE8M0 appears to be more about compatibility than efficiency. DeepSeek hasn't named the source of the chips its new model can use, but the AI startup has reportedly been working closely with Huawei on training and inference using its Ascend family of neural processing units (NPUs). Huawei's Ascend 910C, which powers its CloudMatrix rack systems we looked at last month, doesn't support FP8 natively, suggesting the IT giant may have even more powerful accelerators on the way. Last week, it was reported that DeepSeek had attempted to train its next-gen R2 model on Huawei's Ascend accelerators but struggled to make them work and reverted to using Nvidia H20 accelerators. DeepSeek is now said to be evaluating Huawei's accelerators for inference duty. It's not clear whether or not the so-called R2 refers to the V3.1 model released this week or a forthcoming model. DeepSeek V3.1 isn't really a new model. It was trained from an earlier V3 checkpoint. Despite this, the LLM does promise notable improvements. With V3.1, DeepSeek is no longer differentiating between its "thinking" and "non-thinking" models. V3.1 supports both paradigms in a single model and uses a pair of chat templates to toggle between the two. As such, the company's chatbot interface now omits any reference to R1. The idea of a unified model capable of reasoning and non reasoning outputs isn't new. Alibaba attempted something like this earlier this year but abandoned the idea after finding the functionality degraded the quality of its Qwen 3 models. At least in benchmarking, DeepSeek's V3.1 appears to have avoided that problem. Compared to V3, the point release's non-thinking model achieved significant gains across the board. With thinking enabled, the model's gains were more modest. However that doesn't quite tell the full story, as DeepSeek notes that the model now requires far fewer thinking tokens to arrive at an answer than before, which should help to cut costs associated with serving the model. Speaking of tokens, DeepSeek has boosted the number of tokens in its context window, which you can think of as its short-term memory, from 65,536 to 131,072. While a significant improvement, that still trails other Chinese models like Qwen3, which can handle million-token contexts. DeepSeek also boasted of significant gains in tool and function calling capabilities crucial for agentic AI workloads where external tools and data must be retrieved on the fly. For example, in Browsecomp, a benchmark aimed at autonomous browser use tasks, DeepSeek v3.1 achieved a score of 30 where the May refresh of R1 managed a score of 8.9. Along with access via its Chatbot service and API endpoint, DeepSeek has also made the mode weights for both the base and instruct-tuned models available for download on Hugging Face and ModeScope. ®

China's DeepSeek Release V3.1, Boosting AI Model's Capabilities

DeepSeek announced what appeared to be an update to its older V3 artificial intelligence model on Tuesday, declaring an enhanced version ready for testing. The V3.1 has a longer context window, according to a DeepSeek post to its official WeChat group, meaning it can consider a larger amount of information for any given query. That could allow it to maintain longer conversations with better recall, for example. The Hangzhou-based startup didn't offer much more detail on the update and hasn't posted documentation to major platforms including Hugging Face.

Chinese startup DeepSeek releases upgraded AI model

BEIJING, Aug 21 (Reuters) - Chinese artificial intelligence startup DeepSeek on Thursday released DeepSeek-V3.1, an upgraded model with hybrid inference structure, faster thinking speed and stronger agent capability, the company said in a statement published on WeChat. The company will also adjust the costs for using the model's API, a platform that allows developers of other apps and web products to integrate its AI models, starting September 6, the statement showed. Reporting by Beijing Newsroom; Editing by Christopher Cushing Our Standards: The Thomson Reuters Trust Principles., opens new tab

Chinese AI startup DeepSeek releases upgraded model with domestic chip support

BEIJING, Aug 21 (Reuters) - Chinese artificial intelligence startup DeepSeek released on Thursday an upgrade to its flagship V3 model that the company says has a feature that can optimize it for Chinese-made chips, along with faster processing speeds. The focus on domestic chip compatibility may signal that DeepSeek's AI models are being positioned to work with China's emerging semiconductor ecosystem as Beijing pushes to replace U.S. technology in the face of Washington's export restrictions. DeepSeek shook the technology world this year when it released AI models that compete with Western ones like OpenAI's ChatGPT while offering lower operational costs. The upgrade to DeepSeek's V3 model follows two other recent updates to its core models - an R1 model update in May and an earlier V3 enhancement in March. For domestic chip support, DeepSeek said in a WeChat post its DeepSeek-V3.1 model's UE8M0 FP8 precision format is optimised for "soon-to-be-released next-generation domestic chips". The company did not identify which specific chip models or manufacturers would be supported. FP8, or 8-bit floating point, is a data processing format that allows AI models to operate more efficiently, using less memory while running faster than traditional methods. The DeepSeek-V3.1 features a hybrid inference structure that enables the model to operate in both reasoning and non-reasoning modes, the company said in a WeChat post on Thursday. Users can toggle between these modes using a "deep thinking" button on the company's official app and web platform, both of which now run the V3.1 version. The company will also adjust the costs for using the model's API, a platform that allows developers of other apps and web products to integrate its AI models, starting September 6, the statement showed. Reporting by Beijing Newsroom; Editing by Christopher Cushing and Emelia Sithole-Matarise Our Standards: The Thomson Reuters Trust Principles., opens new tab

DeepSeek hints latest model will be supported by China's 'next generation' homegrown AI chips

Its mention of China's coming next-generation chips may signal plans to work more closely with China's emerging AI chip ecosystem. Chinese artificial intelligence startup DeepSeek has hinted that China will soon have homegrown "next generation" chips to support its AI models, while announcing an update to one of its large language models. In a comment under a post on its official WeChat account, DeepSeek said the "UE8M0 FP8" precision format of its newly released model V3.1 is tailored for the next-generation domestically built chips that will be launched soon. FP8, or 8-bit floating point, is a data processing format that can boost the computational efficiency for training and inference of large deep learning models. DeepSeek's mention of China's coming next-generation chips may signal plans to work more closely with China's emerging AI chip ecosystem in the face of Washington's advanced semiconductor export restrictions and Beijing's push for chip self-sufficiency. The comments come about two weeks after Beijing reportedly urged Chinese AI developers to use domestic alternatives to Nvidia's graphics processing units used in AI training. While analysts say China's domestic AI chipmakers lag behind Nvidia in technological advancement and scale, players like Huawei have been making progress. In its Thursday post, DeepSeek did not disclose the chips it used to train the V3.1, or what local chips the UE8M0 FP8 might be compatible with. DeepSeek shook up the tech world earlier this year after it released its R1 reasoning model, which demonstrated capabilities comparable to those of Western competitors like OpenAI, despite U.S. export controls restricting it from using Nvidia's most advanced AI training chips. Prior to that, in December, the company released its V3 model, which it said had been trained on about 2,000 of Nvidia's less advanced chips. Following DeepSeek's model breakthroughs, the U.S. further tightened export restrictions in April, effectively banning Nvidia's H20 chips, which had been specially designed to meet prior export restrictions on China. Last month, officials from the Trump administration said they planned to allow Nvidia to resume shipping the chips to China. However, the H20s are now being met with scrutiny in China, with regulators reportedly mandating companies against buying the chips until a national security review is completed. Chip analysts have told CNBC that companies like Huawei that have been seeking to build an alternative AI chip ecosystem in China could benefit from a lack of Nvidia's H20s in the market. DeepSeek said Thursday that its V3.1 came with "major changes," including faster response times, and a hybrid reasoning architecture that allows the model to support both reasoning and non-reasoning modes. Reasoning models can execute more complicated tasks through a step-by-step logical thought process. Starting Sept. 6, the company will also adjust the pricing for using the model's API, which allows developers of other apps and web products to integrate DeepSeek on their platforms.

China's DeepSeek just dropped a new GPT-5 rival -- optimized for Chinese chips, priced to undercut OpenAI

DeepSeek's new V3.1 model was quietly released in a message to one of its WeChat groups, China's all-in-one messaging and social app, as well as on the Hugging Face platform. Its debut touches several of today's biggest AI narratives at once. DeepSeek is a core part of China's broader push to develop, deploy, and control advanced AI systems without relying on foreign technology. (And in fact, DeepSeek's new V3 model is specifically tuned to do perform well on Chinese-made chips.) While U.S. companies have been hesitant to embrace DeepSeek's models, they've been widely adopted in China and increasingly in other parts of the world. Even some American firms have built applications on DeepSeek's R1 reasoning model. At the same time, researchers warn that the models' outputs often hew closely to Chinese Communist Party-approved narratives -- raising questions about their neutrality and trustworthiness. China's AI push goes beyond DeepSeek: Its industry also includes models including Alibaba's Qwen, Moonshot AI's Kimi, and Baidu's Ernie. DeepSeek's new release, however, coming just after OpenAI's GPT-5 -- a rollout that fell short of industry watchers' high expectations -- underscores Beijing's determination to keep pace with, or even leapfrog, top U.S. labs. DeepSeek's efforts are certainly keeping U.S. labs on their toes. In a recent dinner with reporters, OpenAI CEO Sam Altman said that rising competition from Chinese open-source models, including DeepSeek, influenced his company's decision to release its own open-weight models two weeks ago. "It was clear that if we didn't do it, the world was gonna be mostly built on Chinese open source models," Altman said. "That was a factor in our decision, for sure. Wasn't the only one, but that loomed large." In addition, last week the U.S. granted Nvidia and AMD licenses to export China-specific AI chips -- including Nvidia's H20 -- but only if they agree to hand over 15% of revenue from those sales to Washington. Beijing quickly pushed back, moving to restrict purchases of Nvidia chips after Commerce Secretary Howard Lutnick told CNBC on July 15: "We don't sell them our best stuff, not our second-best stuff, not even our third-best." By optimizing DeepSeek for Chinese-made chips, the company is signaling resilience against U.S. export controls and a drive to reduce reliance on Nvidia. In DeepSeek's WeChat post, it noted that the new model format is optimised for "soon-to-be-released next-generation domestic chips." Altman, at that same dinner, warned that the U.S. may be underestimating the complexity and seriousness of China's progress in AI -- and said export controls alone likely aren't a reliable solution. "I'm worried about China," he said. Technically, what makes the new DeepSeek model notable is how it was built, with a few advances that would be invisible to consumers. But for developers, these innovations make V3.1 cheaper to run and more versatile than many closed and more expensive rival models. For instance, V3.1 is huge - 685 billion parameters, which is on the level of many top "frontier" models. But its "mixture-of-experts" design means only a fraction of the model activates when answering any query, keeping computing costs lower for developers. And unlike earlier DeepSeek models that split tasks that could be answered instantly based on the model's pre-training from those that required step-by-step reasoning, V3.1 combines both fast answers and reasoning in one system. GPT-5, as well as the most recent models from Anthropic and Google, have a similar ability. But few open weight models have been able to do this so far. V3.1's hybrid architecture is "the biggest feature by far," Ben Dickson, a tech analyst and founder of the TechTalks blog, told Fortune. Others point out that while this DeepSeek model is less of a leap than the company's R1 model -- which was a reasoning model distilled down from the original V3 that shocked the world in January, the new V3.1 is still striking. "It is pretty impressive that they continue making non-marginal improvements," said William Falcon, founder and CEO of AI developer platform Lightning AI. But he added that he would expect OpenAI to respond if its own open source model "starts to meaningfully lag," and pointed out that the DeepSeek model is harder for developers to get into production, while OpenAI's version is fairly easy to deploy. For all the technical details, though, DeepSeek's latest release highlights the fact that AI is increasingly seen as part of a simmering technological cold war between the US and China. With that in mind, if Chinese companies can build better AI models for what they claim is a fraction of the cost, U.S. competitors have reason to worry about staying ahead.

DeepSeek-V3.1 Is 2x Cheaper than GPT-5 | AIM

The hybrid reasoning model comes with improved performance on agentic tasks, says DeepSeek. DeepSeek, the Chinese AI lab backed by High Flyer Capital, has announced a new AI model, called DeepSeek-V3.1. It is a hybrid model that supports both thinking (reasoning) and non-thinking modes. DeepSeek, in the announcement, also called it the first step towards the agent era. The new models improve upon DeepSeek-V3 by providing better tool calling and enhanced thinking efficiency. They deliver answer quality comparable to the reasoning model DeepSeek-R1 while responding quickly. DeepSeek also stated that it has stronger skills in agent-based tasks, and post-training has increased tool use and multi-step task performance. The model has a total of 671 billion parameters, with 37 billion active parameters. DeepSeek's new models are available under the MIT License, and t

Chinese startup DeepSeek releases upgraded AI model - The Economic Times

Chinese artificial intelligence startup DeepSeek released on Thursday an upgrade to its flagship V3 model that the company says has a feature that can optimize it for Chinese-made chips, along with faster processing speeds. The focus on domestic chip compatibility may signal that DeepSeek's AI models are being positioned to work with China's emerging semiconductor ecosystem as Beijing pushes to replace US technology in the face of Washington's export restrictions. DeepSeek shook the technology world this year when it released AI models that compete with Western ones like OpenAI's ChatGPT while offering lower operational costs. The upgrade to DeepSeek's V3 model follows two other recent updates to its core models - an R1 model update in May and an earlier V3 enhancement in March. For domestic chip support, DeepSeek said in a WeChat post its DeepSeek-V3.1 model's UE8M0 FP8 precision format is optimised for "soon-to-be-released next-generation domestic chips". The company did not identify which specific chip models or manufacturers would be supported. FP8, or 8-bit floating point, is a data processing format that allows AI models to operate more efficiently, using less memory while running faster than traditional methods. The DeepSeek-V3.1 features a hybrid inference structure that enables the model to operate in both reasoning and non-reasoning modes, the company said in a WeChat post on Thursday. Users can toggle between these modes using a "deep thinking" button on the company's official app and web platform, both of which now run the V3.1 version. The company will also adjust the costs for using the model's API, a platform that allows developers of other apps and web products to integrate its AI models, starting September 6, the statement showed.

DeepSeek Unveils Upgrade Of Flagship AI Model With Support For Chinese Chips As Beijing Races To Cut Reliance On Nvidia, US Tech - NVIDIA (NASDAQ:NVDA)

On Thursday, Chinese AI startup DeepSeek released an upgraded version of its flagship V3 model, touting support for domestic chips as China accelerates efforts to reduce dependence on U.S. companies like Nvidia Corporation NVDA. DeepSeek's New V3.1 With Domestic Chip Compatibility DeepSeek said its V3.1 model upgrade features faster processing and a new UE8M0 FP8 precision format optimized for "soon-to-be-released next-generation domestic chips," reported Reuters. The company did not disclose which manufacturers or chip models would be supported. FP8, or 8-bit floating point, enables AI models to run more efficiently by using less memory and delivering quicker performance. DeepSeek described the V3.1 update as incorporating a hybrid inference structure, allowing it to toggle between reasoning and non-reasoning modes. A new "deep thinking" button has been added to its app and web platform, now running the upgraded model. DeepSeek also said it will adjust pricing for developers using its API -- which integrates its models into third-party applications -- starting Sept. 6. Also Read: Alibaba Chair Says AI Total Addressable Market Is $10 Trillion, Expects Benefits Via Its Cloud Business Past Struggles With Huawei Hardware Earlier this month, it was reported that DeepSeek delayed its new R2 model launch due to persistent technical problems with Huawei Technologies's Ascend processors. The startup was forced to fall back on Nvidia chips for training, while relying on Huawei hardware only for inference tasks. DeepSeek's Growing Impact On Global AI DeepSeek first made headlines in January when its R1 model triggered a $600 billion sell-off in Nvidia's market value. Founder Liang Wenfeng has since pushed for aggressive model updates, including the V3 release in March and the R1 update in May. Nvidia acknowledged DeepSeek's early success, saying the startup's work demonstrated how models can scale using widely available compute while remaining compliant with U.S. export controls. China's Broader Push For Tech Independence The latest development comes as Beijing pushes for domestic chip adoption. Earlier this week, it was reported that Nvidia has reportedly instructed suppliers to pause production of its H20 AI chip for China, amid pressure from Beijing regulators. Read Next: AMD CEO Lisa Su Says China Strategy Rebounding As MI308 AI Chips Await US License: 'Better Position Than We Were Ninety Days Ago' Photo: mundissima on Shutterstock.com Disclaimer: This content was partially produced with the help of AI tools and was reviewed and published by Benzinga editors. NVDANVIDIA Corp$174.80-0.34%Stock Score Locked: Edge Members Only Benzinga Rankings give you vital metrics on any stock - anytime. Unlock RankingsEdge RankingsMomentum85.52Growth99.28QualityN/AValue6.51Price TrendShortMediumLongOverviewMarket News and Data brought to you by Benzinga APIs

DeepSeek V3.1 Is Here, But It's No Match for GPT-5 or Claude Opus

In the SWE-bench Verified benchmark, the new DeepSeek V3.1 model achieved a score of 66%. DeepSeek, the Chinese AI startup, has launched its new hybrid reasoning model called DeepSeek V3.1 which is designed for agentic use cases and tool calling. It comes with two modes: Think and Non-Think, and can automatically think for longer if the query requires more time to solve. The Think/Non-Think mode can be enabled using the "DeepThink" button. The non-think mode uses deepseek-chat, and the thinking mode uses deepseek-reasoner. Both come with a context length of 128K tokens and activate 37B parameters, out of 671B parameters. For your information, the DeepSeek V3.1 Base is trained on 840B tokens, on top of V3. What is interesting is that DeepSeek V3.1 performs very well at multi-step reasoning tasks. For instance, in SWE-bench Verified -- a benchmark that tests coding performance on real-world software engineering tasks -- DeepSeek V3.1 scored 66.0%, much higher than DeepSeek R1-0528 which got 44.6%. For reference, OpenAI's GPT-5 Thinking scored 74.9% and Anthropic's Claude Opus 4.1 achieved 74.5%. In Humanity's Last Exam (HLE), DeepSeek V3.1 achieved 29.8% with tool calling, and in GPQA Diamond, the new V3.1 model scored 81%. Overall, it seems the new DeepSeek V3.1 model is better than its earlier R1-0528 AI model. However, it doesn't outperform GPT-5 or Claude 4 models. As for API pricing, the DeepSeek V3.1 costs $0.56 / $1.68 for input/output per 1 million tokens.

DeepSeek Marks Push Toward Agentic AI With New Model | 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. The launch of the V3.1 helps keep the Chinese company in the AI game as it prepares its latest iteration of its flagship model, Bloomberg reported Thursday (Aug. 21). V3.1 returns answers to queries faster and is the first step to creating an AI agent, the company said, per the report. The version has been customized to work with next-generation Chinese-made AI chips. DeepSeek rocked the tech world earlier this year when it debuted R1, showing how Chinese companies could take on high-profile rivals in the United States like OpenAI without the most cutting-edge semiconductors. The company's family of models includes V3 (AI chat) and R1 (reasoning). DeepSeek claimed to have spent just $5.6 million to train one of its model, less than the $100 million to $1 billion figure cited by Anthropic. The successor to R1 was expected earlier this year, the Bloomberg report said. Chinese media attributed the hold up to DeepSeek's determination to get the product right, while others speculated about training or development-related hiccups. Meanwhile, a gap exists between awareness of agentic AI and interest in using it. The PYMNTS Intelligence report "The Two Faces of AI: Gen AI's Triumph Meets Agentic AI's Caution" found that while nearly all finance chiefs surveyed said they were aware of the technology, just 15% showed any desire to deploy it within their companies. The gap underscored lingering skepticism among business leaders about the maturity and business value of AI agents as they exist now. Although the technology is promising when it comes to automating complex workflows and improving decision-making, many CFOs remain hesitant due to concerns about implementation risks, oversight challenges and a lack of evidence of ROI. "A lot of companies are excited about what agentic AI can do, but not enough are thinking about what it takes to use it safely," James Prolizo, chief information security officer at Sovos, told PYMNTS this month. "These tools are starting to make real decisions, not just automate tasks, and that changes the game." The research also found that building trust depends on the ability to provide user-friendly reports and visualizations that clearly explain the reasons for an AI agent's actions, as well as ways to provide ongoing human supervision and intervention for critical decisions.

DeepSeek Releases New Version of Model Behind Its AI Chatbot | PYMNTS.com

Citing a post on DeepSeek's official WeChat group, Bloomberg reported that DeepSeek V3.1 is ready for testing. The new version has a longer context window, or space for prompting, of 128,000 tokens. That's roughly 96,000 words or about two 200-page English novels. DeepSeek's V3 model caused a stir in January when the startup claimed it only cost $5.6 million to train using about 2,000 of slower Nvidia chips. That's far cheaper than the millions it took to train frontier models from OpenAI, Google, Anthropic and others. The news wiped $600 billion of market value from Nvidia in one day. But governments soon banned the use of the DeepSeek chatbot out of concerns the data would be kept on Chinese servers. While the startup didn't share much more on WeChat, a post on Reddit said the latest version of the chatbot is "very, very verbose," and also observed that the "r1 in the think button" has disappeared, indicating V3.1 could be a mixed reasoning model. R1 is a reasoning model that DeepSeek also developed. It is offered through the three major U.S. hyperscalers AWS, Microsoft Azure and Google Cloud. The cloud providers have said the model is hosted locally so data would not be sent to China. Developers are still waiting for R2, the next model release of R1, according to Bloomberg. Read also: Remember DeepSeek? Many Adopt Its AI Models Despite Security Concerns In the global AI race, only China is able to compete effectively with the U.S., Bloomberg reported. Chinese companies such as Alibaba, DeepSeek and Moonshot have developed AI models that have capabilities approaching the best ones in the U.S. While the U.S. has banked on largely closed, proprietary AI models, China has pushed open-source models that generally are free to download and use. China is sacrificing short-term profits to ensure Chinese AI is adopted globally, according to Bloomberg. China's 14th five-year blueprint for development, released in 2020, favored the open-source approach. Some Chinese artificial intelligence startup managers also believe the fastest way to enter new markets and compete with U.S. models is to offer open AI models.

DeepSeek releases V3.1: Here's what's new

Open-source DeepSeek V3.1 enhances coding, math, and reasoning under MIT License In a quiet yet impactful move, DeepSeek, the Hangzhou-based AI research lab, has unveiled DeepSeek V3.1, an upgraded version of its already impressive V3 large language model. Announced on August 19, 2025, through the company's official WeChat group, this release has sparked excitement among AI enthusiasts and developers, particularly for its enhanced capabilities and expanded context window. While DeepSeek has kept the official announcement understated, early reports and community buzz on platforms like X suggest V3.1 is a significant step forward in the quest for accessible, high-performance AI. Here's a deep dive into what's new with DeepSeek V3.1 and why it matters. One of the standout upgrades in DeepSeek V3.1 is its expanded context window, now doubled to 128,000 tokens for the online model. This matches the context length of the open-source version, allowing the model to process and retain far more information in a single query. For users, this translates to better handling of long-form conversations, complex document analysis, and retrieval-augmented generation (RAG) tasks. Whether you're summarizing lengthy reports or engaging in multi-turn dialogues, V3.1's ability to "remember" more context ensures more coherent and accurate responses. This upgrade alone makes V3.1 a game-changer for enterprise applications and research tasks requiring extensive data processing. DeepSeek V3.1 pushes the boundaries of scale with a reported 685 billion parameters, up from the 671 billion in its predecessor. This increase, combined with support for multiple tensor formats (BF16, F8_E4M3, and F32), enhances the model's ability to tackle complex reasoning tasks while maintaining efficiency. The model continues to leverage its Mixture-of-Experts (MoE) architecture, activating only 37 billion parameters per token, which keeps inference costs low compared to traditional LLMs. This efficiency, paired with greater computational power, positions V3.1 as a formidable competitor to closed-source giants like GPT-4o and Claude 3.5 Sonnet. Early testing, as reported by communities like Zilliz, highlights significant improvements in DeepSeek V3.1's reasoning, coding, and mathematical abilities. The model excels in logic-driven tasks, with one user noting its success in solving complex problems like "a bouncing ball in a rotating shape." Its coding prowess has also seen a boost, with improved accuracy in generating Python and Bash code, achieving a benchmark score of about 60%, several percentage points higher than the original V3. For math, V3.1 builds on the strengths of its predecessor, which outperformed models like Qwen2.5 72B by a 10% margin on benchmarks like AIME and MATH-500. These enhancements make V3.1 a go-to for developers and researchers tackling technical and analytical challenges. DeepSeek continues its mission to democratize AI by releasing V3.1 under the MIT License, making it freely accessible for developers to use, modify, and share. Available for download on Hugging Face, the model's 685 billion parameters are distributed in the efficient Safetensors format, though it's not yet supported by major inference providers like Hugging Face's Transformers. This open-source approach, combined with V3.1's low training cost, built on the same 2.788 million H800 GPU hours as V3, sets it apart in an industry where training costs can soar into the hundreds of millions. DeepSeek's ability to deliver cutting-edge performance at a fraction of the cost continues to challenge industry giants. For developers, V3.1 maintains compatibility with existing API interfaces, meaning no changes are needed to integrate it into current workflows. The model is accessible via DeepSeek's official website, app, and WeChat mini-program, with the same system prompt and a knowledge cutoff of July 2025. While official benchmarks are still forthcoming, speculation on platforms like Reddit suggests V3.1 could serve as the foundation for an upcoming DeepSeek-R2 model, potentially arriving as early as April or May 2025. This anticipated reasoning-focused model could further elevate DeepSeek's standing in the AI race. DeepSeek V3.1's release underscores the company's relentless pursuit of innovation at an accessible price point. With a training cost of just $5.6 million - compared to $100 million for models like GPT-4 - DeepSeek continues to disrupt the AI landscape, earning it the nickname "the Pinduoduo of AI." Its enhanced context window, increased parameter count, and improved reasoning and coding capabilities make it a versatile tool for developers, researchers, and businesses. As DeepSeek pushes the boundaries of open-source AI, V3.1 signals that the gap between open and closed models is narrowing, setting the stage for a more inclusive AI future. Stay tuned for independent benchmark results and further updates as the community dives deeper into DeepSeek V3.1's capabilities. For now, it's clear this "minor upgrade" is anything but minor, it's a bold step forward in the AI revolution.