DeepSeek has officially released DeepSeek V4-Flash, a new lightweight artificial intelligence (AI) model that offers strong performance while costing much less than many leading AI models.
The company has also open-sourced the model, meaning developers, researchers, and businesses can freely download, modify, and use it under the MIT open-source license, even for commercial projects.
According to DeepSeek, V4-Flash performs better than several popular AI models, including GLM 5.2, and comes close to Anthropic’s Claude Opus 4.8 in many benchmark tests—all while costing only a small fraction of the price.
Built with a Massive AI Architecture
DeepSeek V4-Flash is based on a Mixture of Experts (MoE) architecture.
The model contains 284 billion parameters, although only 13 billion parameters are activated each time it processes a request.
This design allows the model to deliver powerful performance while reducing computing costs and improving response speed.
The complete model repository available on Hugging Face is listed at 304 billion parameters because it also includes DSpark, DeepSeek’s speculative decoding system that helps generate responses faster.
Supports One Million Tokens
One of the biggest features of V4-Flash is its one-million-token context window.
This means the AI can understand and remember a huge amount of information during a single conversation or task.
It can work with:
- Long books
- Large legal documents
- Massive software projects
- Research papers
- Thousands of lines of code
without losing track of earlier information.
Better Performance Than Many AI Models
DeepSeek says the new V4-Flash-0731 version performs better than both the earlier V4-Flash Preview and V4-Pro Preview models.
According to the company, it also outperformed GLM 5.2 across all nine published benchmark tests.
The model performed especially well in coding, software engineering, AI agent tasks, and automation benchmarks.
In several tests, V4-Flash came very close to the performance of Claude Opus 4.8, including:
- Terminal Bench
- DeepSWE
- Agents’ Last Exam
- AutomationBench
- DSBench-FullStack
Claude Opus 4.8 still achieved better results in some areas such as:
- NL2Repo
- Cybergym
- Toolathlon
- DSBench-Hard
However, DeepSeek noted that all benchmark results were measured internally using its own testing methods. Independent testing may produce different results.
Costs More Than 100 Times Less Than Claude
One of V4-Flash’s biggest advantages is its low price.
DeepSeek charges:
- $0.14 per million input tokens
- $0.0028 per million cached input tokens
- $0.28 per million output tokens
According to the company, this makes V4-Flash over 100 times cheaper than Claude Fable 5 for many AI tasks.
Compared to DeepSeek’s own V4-Pro model, V4-Flash’s output is also around three times less expensive.
Designed for Developers
The API supports:
- Up to 2,500 simultaneous requests per account
- One-million-token context
- Outputs of up to 384,000 tokens
- Both reasoning and standard response modes
Developers can also use:
- Function calling
- Responses API compatibility
- Codex support
- Thinking mode
- Non-thinking mode
These features make the model useful for software development, AI agents, coding assistants, and automation tools.
Smarter AI Reasoning
V4-Flash includes three reasoning modes:
- Low
- High
- Maximum
The higher reasoning settings allow the AI to spend more time solving difficult problems before producing an answer.
DeepSeek recommends:
- Temperature: 1.0
- Top-p: 0.95 for AI agent tasks
- Top-p: 1.0 for general tasks
Faster Response Generation
The model includes DSpark, a speculative decoding system that predicts multiple words ahead before verifying them.
According to DeepSeek, DSpark improves response speed by 60% to 85% compared to the company’s previous decoding system without reducing overall performance.
Training and Efficiency
DeepSeek trained V4-Flash using more than 32 trillion tokens, making it one of the largest AI training efforts announced so far.
The company says the model is much more efficient than earlier versions.
At a one-million-token context, V4-Flash reportedly uses:
- Only 10% of the computing power
- About 7% of the memory cache
required by the previous DeepSeek V3.2 model.
This allows the AI to process huge amounts of information while using significantly fewer computing resources.
Self-Hosting Requires Powerful Hardware
Although anyone can use the hosted API without special hardware, running the model locally is much more demanding.
DeepSeek says self-hosting requires powerful servers because all of the model’s experts must remain loaded in memory.
For example:
- A reference setup uses four NVIDIA GB300 GPUs.
- The 8-bit GGUF version of the model is about 162GB.
- The 3-bit version is around 103GB.
Because of these large memory requirements, self-hosting is mainly suitable for companies, research labs, and organizations with high-end multi-GPU systems.
Open Source for Everyone
DeepSeek has released V4-Flash under the MIT license, allowing anyone to:
- Download the model.
- Modify it.
- Use it for commercial products.
- Deploy it on private servers.
This makes it one of the most accessible high-performance AI models currently available.
V4-Pro Still Coming
DeepSeek also confirmed that the final version of V4-Pro is still under development.
The company has not announced when it will be released.
A Powerful Yet Affordable AI Model
With its 284-billion-parameter Mixture of Experts architecture, one-million-token context window, low operating costs, strong benchmark performance, and open-source availability, DeepSeek V4-Flash is designed to offer powerful AI capabilities without the high costs of many commercial models.
However, DeepSeek also noted that its benchmark results are based on the company’s own testing. Developers are encouraged to evaluate the model on their own workloads before deciding whether it meets their needs.
