Try In Playground

DeepSeek V4 Flash

DeepSeek V4 Flash is the efficiency-focused member of the MIT-licensed DeepSeek V4 family — 284 billion total parameters with roughly 13 billion active per token, sharing V4 Pro's hybrid Compressed Sparse Attention and Heavily Compressed Attention architecture and the same 164K served context and 384K output window. The smaller active footprint keeps long-context requests fast and inexpensive, making it well-suited to high-throughput agentic pipelines, document-scale analysis, and latency-sensitive production workloads.

Features

On-demand Deployments

On-demand deployments let you run DeepSeek V4 Flash on dedicated GPUs with Sciforium's high-performance serving stack, with high reliability and no rate limits.

Docs

Low-Latency Long Context

Just 13B active parameters keep long-context requests fast and cheap, without giving up the V4 attention architecture.

Docs
Metadata

State

Ready

Type

LLM

Creator

DeepSeek

Hugging Face

deepseek-ai/DeepSeek-V4-Flash

Specification

Param Count

284B (13B active)

FFN Type

Sparse MoE

Quantization

FP8

Hardware

MI355X

Supported Functionality

Fine-tuning

Contact Sales

On-Demand

Supported

Context Length

164K

Embeddings

-

Input Modality

Text

Output Modality

Text

DeepSeek V4 Flash
DeepSeek V4 Pro
Gemma 4 31B
GLM 5.2
Kimi K2.6
Kimi K2.7 Code
Kimi K3
MiniMax H3
MiniMax M3
Qwen3.5
Qwen3.8
DeepSeek V4 Flash
DeepSeek V4 Pro
Gemma 4 31B
GLM 5.2
Kimi K2.6
Kimi K2.7 Code
Kimi K3
MiniMax H3
MiniMax M3
Qwen3.5
Qwen3.8