DeepSeek V4 Flash

[reasoning]

DeepSeek[text][open-weights]

Efficiency-optimized 284B MoE (13B active) with hybrid attention, 1M context, and high/xhigh reasoning — built for fast, high-throughput coding and agent workflows

[interactive architecture ↓]

Input

$0.11

per 1M tokens

Output

$0.22

per 1M tokens

Context

1.0M tokens

Max Output

66K tokens

Released

2026-04-24

Features

[optional-reasoning][tool-calling][long-context-(1m)][cost-effective]

Technical Specifications

LicenseMIT

Benchmarks

BrowseComp73.2%

GPQA Diamond88.1%

Humanity's Last Exam34.8%

LiveCodeBench91.6%

MMLU-Pro86.2%

SWE-bench Multilingual73.3%

SWE-bench Verified79.0%

Terminal-Bench56.9%

The Lab

location: Hangzhou, Chinafounded: 2023backed by: High-Flyer Quant

DeepSeek is a Chinese AI lab founded in 2023 in Hangzhou by Liang Wenfeng. It is funded solely by High-Flyer Quant, the quantitative hedge fund Liang co-founded in 2015, and grew out of GPU capacity originally stockpiled for trading research. The lab publishes frontier open-weights models under permissive licenses.

more about DeepSeek →

Interactive Architecture

Take DeepSeek V4 Flash apart, layer by layer. The 3D teardown below shows its 43 transformer layers (0 dense + 43 MoE), MLA attention, and expert routing, with values from the public model config.

43 layers (0 dense → 43 MoE)

MLA attention · 64Q heads

6+1 of 256 experts active

284B params · 13B active

Token embedding

4096-dim vectors over a 129,280-token vocabulary, same as V4 Pro.

43 layers + Hash-MoE prefix

18 fewer layers than Pro, with mHC residuals. The first 3 layers use static Hash-MoE routing; the rest use learned MoE routing.

Hybrid CSA + HCA attention

Same Compressed Sparse + Heavily Compressed hybrid as Pro, with a 128-token sliding window and 64 heads. The sparse indexer (64 heads, 512 top-k) keeps the 1M context cheap.

MoE with 6 routed experts

256 routed experts, 6 active per token, 1 shared. Flash activates only 13B of 284B parameters per token, roughly a quarter of Pro's active budget.

Output head

Per-head RMSNorm then projection to 129,280 logits. Flash is the cost-optimised V4 tier for high-volume inference where Pro's 49B active is more than needed.

What vocabulary and embedding size does DeepSeek V4 Flash use?+

4096-dim vectors over a 129,280-token vocabulary, same as V4 Pro.

How many transformer layers does DeepSeek V4 Flash have?+

18 fewer layers than Pro, with mHC residuals. The first 3 layers use static Hash-MoE routing; the rest use learned MoE routing.

What is Multi-head Latent Attention (MLA) in DeepSeek V4 Flash?+

Same Compressed Sparse + Heavily Compressed hybrid as Pro, with a 128-token sliding window and 64 heads. The sparse indexer (64 heads, 512 top-k) keeps the 1M context cheap.

How does the Mixture of Experts in DeepSeek V4 Flash work?+

256 routed experts, 6 active per token, 1 shared. Flash activates only 13B of 284B parameters per token, roughly a quarter of Pro's active budget.

How does DeepSeek V4 Flash produce output tokens?+

Per-head RMSNorm then projection to 129,280 logits. Flash is the cost-optimised V4 tier for high-volume inference where Pro's 49B active is more than needed.

scene values sourced from:HuggingFace config HuggingFace config

Model ID: deepseek-v4-flash

The Lab

location: Hangzhou, Chinafounded: 2023backed by: High-Flyer Quant

more about DeepSeek →

Interactive Architecture

Take DeepSeek V4 Flash apart, layer by layer. The 3D teardown below shows its 43 transformer layers (0 dense + 43 MoE), MLA attention, and expert routing, with values from the public model config.

43 layers (0 dense → 43 MoE)

MLA attention · 64Q heads

6+1 of 256 experts active

284B params · 13B active

Token embedding

4096-dim vectors over a 129,280-token vocabulary, same as V4 Pro.

43 layers + Hash-MoE prefix

18 fewer layers than Pro, with mHC residuals. The first 3 layers use static Hash-MoE routing; the rest use learned MoE routing.

Hybrid CSA + HCA attention

Same Compressed Sparse + Heavily Compressed hybrid as Pro, with a 128-token sliding window and 64 heads. The sparse indexer (64 heads, 512 top-k) keeps the 1M context cheap.

MoE with 6 routed experts

256 routed experts, 6 active per token, 1 shared. Flash activates only 13B of 284B parameters per token, roughly a quarter of Pro's active budget.

Output head

Per-head RMSNorm then projection to 129,280 logits. Flash is the cost-optimised V4 tier for high-volume inference where Pro's 49B active is more than needed.

What vocabulary and embedding size does DeepSeek V4 Flash use?+

4096-dim vectors over a 129,280-token vocabulary, same as V4 Pro.

How many transformer layers does DeepSeek V4 Flash have?+

18 fewer layers than Pro, with mHC residuals. The first 3 layers use static Hash-MoE routing; the rest use learned MoE routing.

What is Multi-head Latent Attention (MLA) in DeepSeek V4 Flash?+

Same Compressed Sparse + Heavily Compressed hybrid as Pro, with a 128-token sliding window and 64 heads. The sparse indexer (64 heads, 512 top-k) keeps the 1M context cheap.

How does the Mixture of Experts in DeepSeek V4 Flash work?+

256 routed experts, 6 active per token, 1 shared. Flash activates only 13B of 284B parameters per token, roughly a quarter of Pro's active budget.

How does DeepSeek V4 Flash produce output tokens?+

Per-head RMSNorm then projection to 129,280 logits. Flash is the cost-optimised V4 tier for high-volume inference where Pro's 49B active is more than needed.