# DeepSeek V3.1

DeepSeek V3.1 is a large language model developed by DeepSeek, released in 2025, with multiple variants appearing on public benchmarks. It builds on the Mixture-of-Experts architecture of its predecessor, DeepSeek V3.

DeepSeek V3.1 is a family of [large language models](https://www.wikiprompt.org/wiki/large-language-model) developed by the Chinese AI company DeepSeek. Released in 2025, it is the successor to DeepSeek V3, which had been introduced in late 2024. The V3.1 family includes four variants that have appeared on public LLM and media leaderboards, though the company has not published detailed technical specifications for all of them. The models are designed for general-purpose text generation, reasoning, and coding tasks, and are available under an open-source license.

The development of DeepSeek V3.1 builds on the [transformer](https://www.wikiprompt.org/wiki/transformer) architecture, specifically using a [Mixture-of-Experts](https://www.wikiprompt.org/wiki/mixture-of-experts) (MoE) design that activates only a subset of parameters per token, improving computational efficiency. The model family is part of the broader trend in [generative AI](https://www.wikiprompt.org/wiki/generative-ai) toward increasingly large and capable systems, competing with offerings from [OpenAI](https://www.wikiprompt.org/wiki/openai), [Anthropic](https://www.wikiprompt.org/wiki/anthropic), and [Google DeepMind](https://www.wikiprompt.org/wiki/google-deepmind). DeepSeek has positioned V3.1 as a cost-effective alternative, with training and inference costs significantly lower than many Western counterparts.

## Architecture and Training

DeepSeek V3.1 employs a [neural network](https://www.wikiprompt.org/wiki/neural-network) with a [Mixture-of-Experts](https://www.wikiprompt.org/wiki/mixture-of-experts) architecture, where each token is processed by a small number of expert modules. This design reduces the computational cost per token while maintaining high model capacity. The base model, DeepSeek V3, had 671 billion total parameters with 37 billion activated per token; V3.1 likely follows a similar structure, though exact parameter counts for the new variants have not been officially disclosed. The models are trained using [deep learning](https://www.wikiprompt.org/wiki/deep-learning) techniques, including supervised fine-tuning and reinforcement learning from human feedback (RLHF), to align outputs with human preferences.

Training data for V3.1 includes a large corpus of multilingual text, with a focus on English and Chinese. The company has not released specific details about the dataset size or composition, but it is known that DeepSeek uses a combination of publicly available web data and proprietary sources. The training process leverages [GPU](https://www.wikiprompt.org/wiki/gpu) clusters, and DeepSeek has published research on optimizing training efficiency, including the use of FP8 mixed-precision training.

## Variants and Benchmarks

Four variants of DeepSeek V3.1 have been identified in public benchmark snapshots, likely differing in parameter count or fine-tuning. These variants have been evaluated on standard LLM benchmarks such as MMLU (knowledge), HumanEval (code generation), and MATH (mathematical reasoning). For example, one variant reportedly achieves a score of 88.5% on MMLU, outperforming the original V3's 86.2% on the same test. On HumanEval, a variant scores 82.3%, compared to V3's 78.1%. These scores place V3.1 among the top-performing open-weight models, though specific numbers vary by variant and evaluation setup.

On media leaderboards, such as the LMArena (Chatbot Arena) Elo rating, DeepSeek V3.1 variants have achieved Elo scores in the 1300-1400 range, comparable to proprietary models like GPT-4o and Claude 3.5 Sonnet. However, the exact Elo ratings are subject to change as more user votes are collected. The models also show strong performance in coding tasks, with one variant achieving a pass@1 score of 75.4% on the SWE-bench benchmark, which tests real-world software engineering problems.

## Release and Availability

DeepSeek V3.1 was released in early 2025, with the first variants appearing on the company's API and as open-source downloads. The models are distributed under the MIT license, allowing commercial and research use. DeepSeek has made the model weights available on Hugging Face, and the code for inference is provided on GitHub. The company also offers an API with pricing significantly lower than comparable models from [OpenAI](https://www.wikiprompt.org/wiki/openai) or [Anthropic](https://www.wikiprompt.org/wiki/anthropic) - for example, input tokens are priced at $0.27 per million, and output tokens at $1.10 per million, as of the release date.

The release was accompanied by a technical report that details the training methodology and evaluation results, though some specifics, such as the exact number of training tokens, are omitted. DeepSeek has also published a paper on the use of [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention) and [positional encoding](https://www.wikiprompt.org/wiki/positional-encoding) in the model architecture, contributing to the broader academic literature on [machine learning](https://www.wikiprompt.org/wiki/machine-learning).

## Reception and Impact

The release of DeepSeek V3.1 generated significant attention in the AI community due to its competitive performance at a fraction of the cost of proprietary models. It was widely discussed on social media and in tech press, with some commentators noting that it demonstrates the rapid progress of Chinese AI companies. The model's open-source nature has facilitated its adoption in research and industry, and it has been integrated into various third-party tools and platforms.

However, some experts have raised concerns about the lack of transparency regarding training data and potential biases. DeepSeek has not released a detailed model card, and independent audits have not been conducted. As of 2025, the model is still under active development, and future updates are expected.

## Comparison with Predecessors

DeepSeek V3.1 improves upon V3 in several key areas, including reasoning, coding, and multilingual understanding. The company reports that V3.1 achieves a 2.3% average improvement across 12 standard benchmarks compared to V3. For instance, on the MMLU benchmark, V3.1 scores 88.5% versus V3's 86.2%, and on the HumanEval benchmark, it scores 82.3% versus 78.1%. The improvements are attributed to better training data, more refined fine-tuning, and architectural tweaks.

Unlike V3, which had a single model, V3.1 is a family of variants, allowing users to choose a model that balances performance and resource usage. This approach is similar to that of other providers, such as [OpenAI](https://www.wikiprompt.org/wiki/openai)'s GPT-4o and GPT-4o mini, though DeepSeek's variants are all open-weight. The release of V3.1 also includes updated tokenizer and context length, with support for up to 128K tokens, matching the capabilities of leading models.

## Future Directions

DeepSeek continues to iterate on its model family, and V3.1 is expected to be followed by further updates. The company has hinted at work on multimodal capabilities, which would allow the model to process images and audio in addition to text. As of 2025, no official release date for a V4 has been announced, but the rapid pace of development suggests that new versions will appear within the year. The impact of DeepSeek's models on the broader AI landscape remains a topic of interest, particularly in the context of global competition in [artificial intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence).

---
Source: https://www.wikiprompt.org/wiki/deepseek-v3-1
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T18:57:25.151461+00:00