With very little fanfare, Chinese AI startup DeepSeek released its most ambitious model on Hugging Face, the 685-billion-parameter open weight large language model DeepSeek V3.1.
Despite the quiet launch, DeepSeek-V3.1 has received widespread attention and early tests revealed performance that rivals proprietary systems from American AI giants, reshaping the competitive landscape with its open-source license and challenging the established economics of frontier AI development.
What is DeepSeek-V3.1?
DeepSeek-V3.1 is a massive 685-billion-parameter model, an increase from its 671B predecessor. It features a 128,000-token context window, which puts it at the level of other open models such as OpenAI’s gpt-oss and Google’s Gemma 3 models. To maintain efficiency while handling complex tasks, the model supports multiple tensor formats, including BF16, F8_E4M3, and F32, which allows developers to optimize performance for their specific hardware.
The model is based on the Mixture-of-Experts (MoE) architecture that activates only 37 billion parameters per token, which helps keep inference costs low despite its immense total size. It has a hybrid design that integrates reasoning and non-reasoning functions into a single model. The model has also been trained for native tool use, search, and coding. The model works in normal chat mode by default and can switch to thinking and tool-use by adding special tokens.
This design is contrary to the previous generation of DeepSeek models, which had separate models for normal tasks (DeepSeek-V2) and reasoning (DeepSeek-R1).
DeepSeek has apparently managed to solve challenges that have caused performance issues in previous attempts at hybrid models. (Two things to look out for: 1) What kind of new techniques DeepSeek employed, which is likely to appear in the technical report they will probably release soon, and 2) Whether they have abandoned plans to release DeepSeek-R2 now that they have a hybrid model that also includes reasoning capabilities.)
DeepSeek-V.31 has been released under the permissive MIT license, which means it is available for commercial use and modification. DeepSeek has released the base model and the hybrid model on Hugging Face.
It is also available on DeepSeek’s API at $0.56 per million input tokens and $1.68 per million output tokens.
Performance that rivals the proprietary giants
Early benchmarks show DeepSeek-V3.1 achieving a 71.6% score on the Aider coding benchmark. This result places it slightly ahead of proprietary models like Anthropic’s Claude Opus 4 while being significantly cheaper. According to VentureBeat, the model delivers this performance at roughly $1.01 per complete coding task, compared to systems that cost nearly $70 for an equivalent workload.
Beyond coding, the model shows strong reasoning abilities. Early testers reported its success in solving complex logic problems, such as the famous “bouncing ball in a rotating shape.” Its mathematical skills also build on the strengths of its predecessor, which had already outperformed other models on benchmarks like AIME and MATH-500.
This cost-efficiency stems partly from a low training cost. We still need training figures for the new model. But we know that its predecessor, DeepSeek-V2, required just $5.6 million for one training run, which makes it significantly cheaper than equivalent efforts by U.S. labs.
A new world order in AI development
The global developer community responded immediately, with V3.1 quickly climbing the trending list on Hugging Face. This rapid embrace shows that technical merit is a primary driver of adoption, regardless of a model’s national origin.
The release was also strategically timed, arriving just weeks after OpenAI’s GPT-5 and Anthropic’s Claude 4.1 launches. By nearing their performance with an open-source model, DeepSeek directly challenges the high-cost, closed business models of its American competitors. OpenAI’s CEO, Sam Altman, recently told CNBC that rising competition from Chinese open-source models, including DeepSeek, influenced his company’s decision to release its own open-weight models.
But it is worth noting that while DeepSeek-V3.1 is freely available, its enormous 700GB size presents a significant practical barrier. Self-hosting and customizing the model requires substantial computational resources and expertise that most organizations lack. For many, the main benefit of its “openness” will be access to lower-cost APIs through cloud providers, rather than the ability to run and modify it locally. Furthermore, U.S. enterprises may hesitate to adopt the model due to geopolitical tensions and a preference for domestic vendors that offer integrated platforms and enterprise-grade support and security.
(I would definitely keep an eye on tiny language models that can be easily deployed on a wide range of consumer hardware and single accelerators, such as the Gemma 3 family. For enterprise applications, small models that can be easily customized for spethrough curated datasets or distillation might prove to be more practical than “one model to rule them all” systems.)
A new benchmark for power and access
Just as U.S. AI companies are trying to up their open model game, DeepSeek-V3.1 might set a new benchmark, proving that frontier-level AI performance can be achieved and distributed openly and at a fraction of the traditional cost.
The continued release of state-of-the-art open models can change the dynamics of the global AI race, shifting the focus from who can build the most powerful system to who can make that system the most accessible. Its release, however, also raises an important question for the open-source community: as models grow to this immense scale, the line between a powerful, accessible utility and a truly customizable tool becomes more distinct. The future of AI will be defined by the balance between raw power and practical usability.