DeepSeek V4 Flash: the fast model of the V4 series
Overview
DeepSeek V4 Flash: 284B parameters, 13B active, 1M token context
Flash is the compact and fast variant of the DeepSeek V4 series. With 284B total parameters and 13B activated per token, it retains the same 1M token context window as the Pro model while offering lighter inference. Available on OpenRouter at $0.14/M input tokens and $0.28/M output tokens.
Flash Architecture
284B total parameters, 13B activated per inference
Flash uses DeepSeek's MoE (Mixture of Experts) architecture with hybrid attention and diversity-constrained hyper-connections. Only 13B parameters are activated per token, significantly reducing inference cost compared to Pro (49B active).
Ideal for daily tasks, quick summaries, and high-volume workflows.
1M Token Context
Same context window as Pro: 1 million tokens
Despite its smaller size, Flash retains the 1M token context window. DeepSeek reports that V4-Flash uses only 10% of DeepSeek-V3.2's KV cache in the 1M token scenario, thanks to hybrid attention and architecture optimizations.
Test Flash on long documents before switching to Pro if the task is more complex.
Model choice
Flash vs Pro
Flash: 284B total / 13B active. Pro: 1.6T total / 49B active. Same 1M token context.
Flash is the default free entry point. Pro is reserved for unlimited subscriptions and tasks requiring deeper reasoning.
Technical
MoE Architecture
Mixture of Experts with hybrid attention, diversity-constrained hyper-connections, and Muon optimizer.
The MoE architecture activates only a fraction of parameters per token, enabling a large model while keeping inference cost reasonable.
Usage
Reasoning modes
Non-think, Think High, and Think Max to adjust analysis depth.
Non-think prioritizes speed. Think High improves accuracy. Think Max pushes reasoning to the maximum, recommended with at least 384K tokens of context.
Evaluation
Benchmarks
MMLU-Pro, HumanEval, GSM8K, LongBench-V2, LiveCodeBench, SWE Verified, MCPAtlas.
Official tables cover general knowledge, reasoning, code, math, long context, and agentic tasks.
Integration
OpenAI-compatible API
API identifier: deepseek-v4-flash. OpenAI and Anthropic compatible format.
Use deepseek-v4-flash in your existing API integrations. Recommended temperature: 1.0, top_p: 1.0.
Deployment
Open weights
Weights available on Hugging Face for local or cloud deployment.
Flash can be run locally. The model card includes encoding instructions, sampling settings, and compatibility notes. FP8 supported.
Why Flash
Flash is designed for speed without sacrificing long context
With 13B active parameters and 1M token context, Flash offers a good balance between cost, speed, and capability for everyday tasks.
Lightweight inference
Only 13B parameters activated per token. DeepSeek reports Flash uses 27% of DeepSeek-V3.2's single-token inference FLOPs.
Optimized KV cache
10% of DeepSeek-V3.2's KV cache in the 1M token scenario, thanks to hybrid attention.
Adjustable reasoning
Non-think for maximum speed, Think High for more accuracy, Think Max for difficult tasks.
Code and agents
Evaluated on LiveCodeBench, SWE Verified, Toolathlon, and MCPAtlas for developer and agentic workflows.
Resources
Official DeepSeek V4 Flash links
Access weights, source code, and official documentation to deploy or evaluate Flash.
Weights and model card
- Official model card with benchmarks and deployment instructions.
- Weights available for local and cloud inference.
- FP8 instructions, encoding, and recommended sampling parameters.
Source code
- GitHub repository with integration examples and scripts.
- Compatible with standard inference frameworks.
- Documented prompt examples and use cases.
Recommended usage
- Temperature 1.0, top_p 1.0 for local deployment.
- Minimum 384K tokens of context for Think Max.
- Test your own documents before choosing between Flash and Pro.
Official data
DeepSeek V4 Flash benchmarks: what the numbers say
The official model card publishes results on knowledge, reasoning, code, math, long context, and agentic tasks. Here are the key points.
Compare Flash and Pro on the benchmarks that match your real use cases, not just general rankings.
Flash: 284B total parameters, 13B active. Pro: 1.6T total parameters, 49B active. Same 1M token context.
Benchmarks covered: MMLU-Pro, HumanEval, GSM8K, LongBench-V2, LiveCodeBench, SWE Verified, Toolathlon, MCPAtlas.
Flash uses 27% of single-token inference FLOPs and 10% of DeepSeek-V3.2's KV cache in the 1M token scenario.
Instruct modes: Non-think (speed), Think High (accuracy), Think Max (maximum reasoning, min. 384K tokens).
Speed and cost
Flash for fast tasks and high-volume workflows
With 13B active parameters and a cost of $0.14/M input tokens, Flash is the natural choice for daily use, summaries, and high-throughput API integrations.
- Document summaries, emails, everyday writing.
- API integrations with high request volume.
- Quick comparison of multiple responses before switching to Pro.
Long context
1M token context even on the Flash model
Flash retains the same context window as Pro. Test it on your long documents, codebases, or multi-step analyses before deciding if Pro is necessary.
- Contracts, manuals, long technical documentation.
- Large codebases for review or refactoring.
- Multi-layer analyses in a single context.
Local deployment
Deploy Flash locally or via API
Flash's open weights are available on Hugging Face. The model card includes encoding instructions, recommended sampling parameters, and compatibility notes.
- Weights available on HuggingFace for local deployment.
- FP8 supported to reduce memory footprint.
- Compatible with standard inference frameworks.
FAQ
DeepSeek V4 Flash: basics and architecture
Answers to the most common questions about the Flash model.
Flash is the compact variant of the DeepSeek V4 series. 284B total parameters, 13B activated per token, 1M token context. It's the default free entry point.
Pro: 1.6T total parameters, 49B active. Flash: 284B total, 13B active. Both have 1M token context. Pro is more powerful, Flash is faster and cheaper.
MoE (Mixture of Experts) with hybrid attention, diversity-constrained hyper-connections, and Muon optimizer. Same architectural family as Pro.
Yes, weights are available on Hugging Face. Check the license on the official model card.
FAQ
Performance and reasoning modes
What benchmarks and instruct modes mean in practice.
MMLU-Pro, HumanEval, GSM8K, LongBench-V2, LiveCodeBench, SWE Verified, Toolathlon, MCPAtlas. Cover knowledge, code, math, long context, and agents.
Non-think: fast response without extended reasoning. Think High: more accuracy. Think Max: maximum reasoning, requires at least 384K tokens of context.
Yes. DeepSeek reports Flash uses 10% of V3.2's KV cache in the 1M token scenario thanks to hybrid attention.
No. Flash is more compact. For complex tasks, Pro remains more suitable. Test your own workflows to decide.
FAQ
Deployment, API, and resources
How to use Flash in production or locally.
API identifier: deepseek-v4-flash. OpenAI and Anthropic compatible format. Available on OpenRouter at $0.14/M input tokens.
Temperature 1.0, top_p 1.0 for local deployment according to the official model card.
Weights are available on Hugging Face. FP8 supported to reduce memory footprint.
The GitHub repository contains integration scripts, prompt examples, and technical documentation.
Resources
Everything you need to know about DeepSeek V4 Flash
Architecture, benchmarks, reasoning modes, API, local deployment, and comparison with Pro.
Get started
Test DeepSeek V4 Flash on a real task
Start with a summary, code review, or long document. Compare Flash and Pro on the same workflow to choose the right model.