Model library / deepseek-ai
DeepSeek-V4-Flash-0731 is a 304.2 billion-parameter, mixture-of-experts language model with speculative decoding for agentic tasks.
6 of 256 experts active per token
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
DeepSeek-V4-Flash-0731 is released by deepseek-ai as the official version of DeepSeek-V4-Flash, extending the preview model with enhanced agentic capabilities. The publisher describes it as suitable for code-agent tasks and general reasoning, offering a speculative decoding module (DSpark) and configurable reasoning effort levels. It is intended for use with OpenAI-compatible chat encoding and supports tool calling and reasoning features.
In practice the model comprises 304.2 billion parameters, organised as 256 experts with six active per token, and operates in fp8 format. It supports a context window of 1,048,576 tokens and is distributed under an MIT licence. Deployment requires handling the large size, using vLLM or SGLang with DSpark speculative decoding, and configuring sampling parameters such as temperature 1.0 and top_p 0.95.
Specification
| Repository | deepseek-ai/DeepSeek-V4-Flash-0731 |
| Publisher | deepseek-ai |
| Published | 31 July 2026 |
| Architecture | DeepseekV4ForCausalLM |
| Model type | deepseek_v4 |
| Parameters | 304,180,418,494 |
| Experts | 256 total, 6 active per token |
| Layers | 43 |
| Hidden size | 4,096 |
| Attention heads | 64, 1 key/value heads |
| Context length | 1,048,576 tokens |
| Vocabulary | 129,280 tokens |
| Quantisation | fp8 |
| Licence | mit |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 566.6 GiB | Full precision as released |
| fp8 | 283.3 GiB | 8-bit, near-lossless on most models |
| int4 | 141.6 GiB | 4-bit, smallest footprint |
The cache costs 86 KiB per token at 16-bit, so the full 1,048,576-token context of one request needs about 86.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 128-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | more than 8 | — | — |
| RTX 5090 32GB | 8 | 55.0 GiB | about 670,388 tokens |
| L40S 48GB | 4 | 7.4 GiB | about 90,014 tokens |
| A100 80GB | 4 | 122.6 GiB | the full 1,048,576 tokens |
| H100 80GB | 4 | 122.6 GiB | the full 1,048,576 tokens |
| RTX PRO 6000 96GB | 2 | 12.4 GiB | about 150,977 tokens |
| H200 141GB | 2 | 93.4 GiB | the full 1,048,576 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \ --tensor-parallel-size 8 \ --max-model-len 667648 \ --gpu-memory-utilization 0.90
Questions
The weights take about 155 GiB at the released precision. No single card in the table above holds that, so it needs several GPUs or a lower precision.
The repository declares mit. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 1,048,576 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
DeepSeek-V4-Pro-0813 is a 1650.5B mixture-of-experts open-weight model published by deepseek-ai on Hugging Face. It uses the Deep…
Nex-N2.5-Pro is a 396.8 billion-parameter MoE language model designed for long-horizon, agentic tasks with multimodal and compute…
GigaChat 3.5 Reasoning is a 438 billion-parameter mixture-of-experts language model trained with online RL for improved reasoning…
K2-Horizon-375B-A23B is an open-weight sparse MiE language model from IFM with 379 B parameters, 23 B active per token and a 512 …
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.