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-V4-Flash

DeepSeek-V4-Flash

Efficient mixture-of-experts model with one million token context

DeepSeek-V4-Flash is a large language model released by DeepSeek AI in 2026. It uses a mixture of experts architecture to activate only 13 billion parameters from a total of 284 billion. The model supports a context window of one million tokens and offers open weights under an MIT licence.

Share on WhatsApp

  • Activates only 13 billion of 284 billion parameters
  • Supports a context window of one million tokens
  • Released under an open MIT License
  • Uses hybrid attention for long-context efficiency
  • Offers three distinct reasoning effort modes

What it is

DeepSeek-V4-Flash is a general-purpose language model developed by DeepSeek AI. It belongs to the DeepSeek-V4 series and was released as a preview before an official update in July 2026. The model contains 284 billion parameters but activates only 13 billion during inference, making it highly efficient. It supports a context length of one million tokens, allowing it to process vast amounts of text at once. The model is designed for tasks requiring strong reasoning and agentic capabilities.

How it came about

The model was introduced as part of the DeepSeek-V4 series, which aims for efficient intelligence over long contexts. A preview version was released first, followed by the official DeepSeek-V4-Flash-0731 release. This later version superseded the preview and included enhanced agentic capabilities. It shares its structure with the DSpark variant, which includes a speculative decoding module. The development involved pre-training on more than 32 trillion tokens and a two-stage post-training pipeline.

Inside the model

The architecture uses a Mixture of Experts design with hybrid attention mechanisms. These include Compressed Sparse Attention and Heavily Compressed Attention to improve efficiency for long contexts. Manifold-Constrained Hyper-Connections strengthen signal propagation across layers. The model was trained using the Muon optimizer for stability. Post-training involved independent cultivation of domain experts via supervised fine-tuning and reinforcement learning, followed by unified consolidation through distillation.

What it can do

DeepSeek-V4-Flash performs strongly on knowledge and reasoning benchmarks. It achieves high scores on MMLU-Pro and coding tests like LiveCodeBench. The model supports three reasoning effort modes: non-think for fast responses, think high for logical analysis, and think max for complex problem-solving. In its official release, it outperformed the preview Pro version on several agentic benchmarks such as Terminal Bench and Toolathlon. It handles multilingual tasks effectively, including Chinese and English evaluations.

How to use it

The model weights are available openly under the MIT License. Users can download them from Hugging Face and ModelScope. The official release supports FP4 and FP8 mixed precision. It is compatible with inference engines like vLLM and SGLang, which support its speculative decoding features. The company provides Python scripts for encoding messages in an OpenAI-compatible format, as no Jinja chat template is included.

Questions people ask about DeepSeek-V4-Flash

How many parameters does DeepSeek-V4-Flash have?

The model has a total of 284 billion parameters. However, it uses a mixture of experts architecture that activates only 13 billion parameters for each token during inference.

What is the maximum context length?

DeepSeek-V4-Flash supports a context length of one million tokens. This allows it to process and reason over extremely large documents or codebases in a single pass.

Is the model open source?

Yes, the model weights are released under the MIT License. They are available for download on platforms like Hugging Face and ModelScope, allowing for local deployment and modification.

What reasoning modes are available?

The model supports three reasoning effort levels: low for fast intuitive answers, high for conscious logical analysis, and max for pushing reasoning capabilities to their limit on complex tasks.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on DeepSeek-V4-Flash yet, so the facts on this page come from the sources listed above: huggingface.co/deepseek-ai/DeepSeek-V4-Flash, huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731. 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.