DeepSeek V3.5
DeepSeek V3.5 is a mixture-of-experts model with 685B total parameters and 37B activated per token. Notable for delivering frontier performance at a fraction of the training cost of competing models.
Released
March 24, 2026
Type
llm
License
open-weight
Context
128,000 tokens
Pricing
Input
$0.27 / 1M tokens
Output
$1.10 / 1M tokens
Capabilities
Architecture
Parameters: 685B MoE (37B active)
Links
DeepSeek V3.5 in the news
Bloomberg Markets · Aug 3, 2026
New DeepSeek, Alibaba Models Take On Anthropic, OpenAI | The China Show | 8/3/2026
Chinese artificial intelligence company DeepSeek has launched its latest large language model, DeepSeek-V2, aiming to compete with leading global AI developers. This release, detailed in a company blog post on April 15, 2024, signifies a significant advancement in China's AI capabilities. DeepSeek-V2 reportedly achieves performance comparable to models like Anthropic's Claude 3 Opus and OpenAI's GPT-4 while utilizing substantially fewer computational resources. Specifically, DeepSeek-V2 is claimed to require only 20% of the GPU memory needed by comparable models, a critical factor for efficient deployment and scaling. The model's architecture incorporates a novel Mixture-of-Experts (MoE) approach, which allows for greater efficiency by activating only relevant parts of the neural network for specific tasks. This innovation is key to its reduced memory footprint and enhanced processing speed. Alibaba Cloud, the cloud computing arm of Chinese e-commerce giant Alibaba, has also introduced its new AI model, Qwen1.5. Announced on April 18, 2024, Qwen1.5 is available in various sizes, with the largest version boasting 110 billion parameters. This model demonstrates strong performance across a range of benchmarks, including coding, mathematics, and general knowledge, positioning it as a formidable competitor in the rapidly evolving AI landscape. Alibaba's move underscores the intense competition within the AI sector, both domestically in China and on the global stage. The development and release of these advanced models by Chinese companies reflect the nation's strategic focus on becoming a leader in artificial intelligence technology. The availability of these powerful models is expected to drive innovation and adoption of AI solutions across various industries within China and potentially beyond. The competitive landscape for large language models is increasingly crowded, with companies worldwide investing heavily in research and development. Anthropic, known for its Claude series of models, and OpenAI, the creator of the GPT series, have set high benchmarks for performance and capability. DeepSeek-V2's claim of superior efficiency, requiring significantly less GPU memory, addresses a major bottleneck in deploying large AI models. This could make advanced AI more accessible and cost-effective for a wider range of businesses and applications. The MoE architecture, a key feature of DeepSeek-V2, is a growing trend in AI research, enabling models to scale more effectively without a proportional increase in computational cost. The success of this approach in DeepSeek-V2 could influence future model development across the industry. Alibaba's Qwen1.5, with its substantial parameter count and broad capabilities, further intensifies this competition. The model's performance on benchmarks suggests it can handle complex tasks, making it a valuable tool for developers and enterprises. The release of Qwen1.5 by Alibaba Cloud also highlights the strategic importance of AI for major technology companies seeking to expand their cloud services and offer cutting-edge AI solutions to their customers. As these new models become more widely available and tested, their impact on the global AI market will become clearer, potentially shifting the balance of power among leading AI developers and fostering new waves of AI-driven innovation.
Fortune · Aug 1, 2026
Despite China’s 996 culture, DeepSeek founder Liang Wenfeng says his workers don’t do overtime or even have KPIs: ‘No one manages them’
In a remarkable departure from China's notoriously demanding tech industry culture, DeepSeek, an artificial intelligence company founded by Liang Wenfeng, has achieved a $60 billion valuation in a mere three years, positioning itself as a significant rival to established players like OpenAI's ChatGPT. Since its public launch in January 2025, DeepSeek's AI models have garnered approximately 173 million downloads, underscoring its rapid market penetration and technological prowess. However, the company's operational philosophy, as articulated by its founder, stands in stark contrast to the prevalent "996" work culture – a schedule mandating 9 a.m. to 9 p.m. workdays, six days a week, which was officially outlawed in China in 2021 but has persisted as a de facto standard in many tech firms. Liang Wenfeng, whose personal net worth is estimated at $37.9 billion, has revealed that DeepSeek employees generally do not work overtime and operate without the imposition of Key Performance Indicators (KPIs). He attributes this unconventional approach to two core principles: the belief that research necessitates a relaxed and unhurried environment, arguing that excessive pressure can stifle innovation, and a deliberate focus on undertaking very few, high-impact tasks. This philosophy permeates the company's structure, which reportedly eschews rigid hierarchies, formal rulebooks, and strict, daily written objectives. Instead, DeepSeek is described as being "driven by a vision, organized by a vision," a vision that Liang states is not even formally documented. Employees at DeepSeek are afforded a significant degree of autonomy, with Liang indicating that "formal work" is not expected to consume more than half of an employee's day. The remaining half is intentionally unallocated, allowing individuals the freedom to pursue their own interests or projects, fostering an environment of intrinsic motivation and self-direction. This management style is particularly noteworthy given the intense competitive landscape of the global AI sector, where companies are often under immense pressure to deliver rapid advancements. DeepSeek's success, evidenced by its substantial valuation and widespread adoption of its products, suggests that alternative, more employee-centric management models can indeed yield exceptional results, challenging the traditional paradigms of high-pressure tech environments.
MarkTechPost · Jul 31, 2026
DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains
DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and simultaneously moved the official V4-Flash API into public beta on July 31, 2026. This release represents the official version, superseding a previous preview, with no changes to the underlying architecture or model size. The performance improvements stem from re-post-training rather than a new design. The model checkpoint includes the DSpark speculative decoding module, aligning with the structure of DeepSeek-V4-Flash-DSpark. Hugging Face lists the repository's parameter count at 304 billion, which accounts for the DSpark module layered on top of the 284 billion base parameters. On the API front, the deepseek-v4-flash model now natively supports the Responses API format and has been adapted for coding tasks, specifically referencing Codex. Notably, the V4-Pro API, along with the associated app and web models, did not receive updates in this release. Deployment of DeepSeek-V4-Flash-0731 is feasible through two distinct methods. Via API, DeepSeek offers pricing at $0.14 per 1 million input tokens during a cache miss and $0.0028 per 1 million input tokens on a cache hit. Output tokens are priced at $0.28 per 1 million, with a concurrency limit of 2,500. This API pricing is approximately one-third of the output cost for deepseek-v4-pro, which is listed at $0.87 per 1 million tokens. This pricing structure enables seed-stage startups, independent developers, and internal platform teams to run agent loops without requiring a dedicated GPU budget. Self-hosting presents a higher barrier to entry. While the model weights are available under an MIT license and are ungated, all experts within the model remain resident in memory, even though only 13 billion parameters are activated per token. DeepSeek's provided vLLM example demonstrates serving the model on a single 4xGB300 node. Using Unsloth's dynamic GGUFs, a lossless 8-bit build requires 162 GB of storage, and a 3-bit build requires 103 GB, necessitating approximately 110 GB of combined RAM and VRAM. Self-hosting is therefore best suited for mid-size to large enterprises equipped with a serving cluster or a single, high-performance workstation capable of aggressive quantization. According to the DeepSeek-V4 technical report, V4-Flash is a Mixture-of-Experts (MoE) model with 284 billion parameters. It activates 13 billion parameters per token and features a context window of 1 million tokens. Each MoE layer incorporates one shared expert and 256 routed experts, with an intermediate dimension of 2048. Six routed experts are activated for each token. The initial three MoE layers are detailed in the technical report.
Fortune · Jul 25, 2026
DeepSeek said to tell backers of funding pause after viral posts
DeepSeek, a Chinese artificial intelligence company, has informed prospective investors that it is temporarily suspending its second fundraising round. This decision comes days after comments widely attributed to its founder, Liang Wenfeng, regarding US-Chinese AI competition gained significant traction online. The company verbally communicated to some potential backers that investment agreements would not be finalized as anticipated in the coming days. DeepSeek may consider resuming the fundraising process at a later date, according to sources familiar with the matter who requested anonymity to discuss private deliberations. The suspension appears to be partly driven by Liang's dissatisfaction with online reports concerning his remarks made to investors during the company's first financing deal. That initial round, which closed in June, successfully raised $7 billion for the AI lab. Bloomberg has not independently verified the authenticity of the viral posts, which reportedly included a transcript from a meeting Liang conducted with undisclosed individuals. Chinese media outlets, including Yicai, reported this week that the billionaire founder discussed the reliance on Nvidia Corp. chips for AI development and China's ongoing deficit in AI sophistication compared to the United States. DeepSeek has not yet responded to an emailed request for comment regarding the transcript or the fundraising pause. Negotiations are described as fluid, and the company might still decide to proceed with the deal. It remains unclear whether DeepSeek has informed all potential investors in the current round of its intentions. The current funding talks commenced only weeks after DeepSeek concluded its record-breaking first financing round, which attracted prominent investors such as Tencent Holdings Ltd. and Contemporary Amperex Technology Co. Ltd. Prior to the suspension, DeepSeek was reportedly aiming to secure at least 10 billion yuan in additional funds through this follow-on deal, with the final amount potentially increasing based on investor participation. The startup had been targeting a pre-money valuation of at least 480 billion yuan, according to Bloomberg's previous reporting.
Fortune · Jul 22, 2026
As Washington panics about Chinese AI, Jensen Huang says open-source models like Kimi are ‘excellent’ and should be embraced, not banned
Nvidia CEO Jensen Huang urged Washington to embrace open-source AI models developed by Chinese companies, rather than restricting them due to perceived threats. Speaking to Axios on Tuesday, Huang stated that these highly capable, low-cost models, including Moonshot AI's Kimi K3, DeepSeek, and Alibaba's offerings, are "excellent" and should be utilized. This stance comes as the White House reportedly considers restricting U.S. companies' use of Chinese AI models, driven by concerns over potential surveillance and the impact on domestic AI companies. The White House has allegedly considered executive actions to impose conditions on U.S. firms using these models, such as mandating security guarantees and liability acceptance in case of breaches. Huang dismissed the notion that these models could serve as a backdoor for the Chinese government, calling it a "misconception." He emphasized that these models are downloadable and their security guardrails are customizable. Huang also expressed confidence that these open-source models will not surpass American AI advancements, arguing for the necessity of both open-source and closed-source AI systems. He believes that excellent open-source models should be integrated into the AI ecosystem, alongside proprietary models from companies like OpenAI and Anthropic. U.S. companies, including AI coding startup Cursor, are reportedly turning to these Chinese open-source alternatives to mitigate the high costs associated with American AI models. Nvidia did not immediately respond to requests for comment.