News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
DeepSeek-Coder-V2

DeepSeek-Coder-V2

Open-source mixture-of-experts model for code and math

DeepSeek-Coder-V2 is an open-source code language model released by DeepSeek-AI in June 2024. Built on a mixture-of-experts architecture, it supports 338 programming languages and a 128k context window. The model aims to match closed-source systems in coding and mathematical reasoning tasks.

Share on WhatsApp

  • Supports 338 programming languages
  • Context window extends to 128k tokens
  • Uses mixture-of-experts for efficiency
  • Trained on 6 trillion additional tokens
  • Open weights with commercial licence

What it is

DeepSeek-Coder-V2 is a large language model designed for software engineering and mathematical reasoning. Developed by DeepSeek-AI, it was announced in June 2024. The system uses a mixture-of-experts design, which activates only a subset of its parameters for each token to improve efficiency. It comes in two sizes: a Lite version with 16 billion total parameters and a larger version with 236 billion. Both variants support a context window of 128,000 tokens, allowing them to process extensive codebases or long documents in a single pass.

How it came about

The model builds upon an intermediate checkpoint of DeepSeek-V2. Developers continued pre-training this base with an additional 6 trillion tokens to enhance specific capabilities. This process aimed to improve coding and math skills while preserving general language performance. It succeeds DeepSeek-Coder-33B, offering significant advancements in code-related tasks and reasoning. The release includes both base and instruct versions, making the weights publicly available for research and commercial use.

Inside the model

The architecture relies on the DeepSeekMoE framework. The Lite model contains 16 billion total parameters but activates only 2.4 billion during inference. The full model has 236 billion total parameters with 21 billion active. This sparse activation reduces computational cost. Training involved ingesting 6 trillion extra tokens focused on code and logic. The system supports 338 programming languages, a major increase from the 86 supported by its predecessor. Tokenizers and chat templates are provided for integration with standard libraries like Hugging Face Transformers and vLLM.

What it can do

DeepSeek-AI claims the model performs comparably to GPT-4 Turbo in code-specific tasks. Benchmark evaluations suggest it surpasses closed-source rivals such as Claude 3 Opus and Gemini 1.5 Pro in coding and mathematics. It handles code completion, insertion, and chat-based instructions. The expanded language support allows it to work with a wide variety of legacy and modern programming syntaxes. Mathematical reasoning capabilities were also substantially enhanced through the continued pre-training phase.

How to use it

The model weights are open for public download via Hugging Face. Both base and instruct variants are available. The code repository uses an MIT licence, while the model weights follow a specific Model License that permits commercial use. Developers can run the Lite version locally using GPUs with BF16 precision. An API is also offered through the DeepSeek Platform with pay-as-you-go pricing. A dedicated chat interface is accessible on the official website.

Questions people ask about DeepSeek-Coder-V2

How many parameters does DeepSeek-Coder-V2 have?

The model comes in two sizes. The Lite version has 16 billion total parameters with 2.4 billion active. The full version has 236 billion total parameters with 21 billion active per token.

Can I use DeepSeek-Coder-V2 for commercial projects?

Yes. The series supports commercial use. The code is under an MIT licence, and the model weights are subject to a Model License that allows commercial application.

What is the context length of the model?

Both the Lite and full versions of DeepSeek-Coder-V2 support a context length of 128,000 tokens. This is a significant increase from the 16k limit of previous iterations.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on DeepSeek-Coder-V2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2406.11931, huggingface.co/deepseek-ai/DeepSeek-Coder-V2-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.