DeepSeekOpen Source

DeepSeek V4 Flash

DeepSeek V4 Flash is a reasoning model developed by DeepSeek-AI. It has a large parameter count of approximately 2840B and a long context window of 1M tokens.

Parameters

2840.0B

Context Window

License

MIT

Release Date

2026-04-24

API Pricing

Input Price (per 1M tokens)

$0.14

Output Price (per 1M tokens)

$0.28

Billing Mode: standard

Strengths

・Massive 2840B parameters
・Long-context understanding of 1M tokens
・High freedom via MIT license

Weaknesses

・High computational load during inference
・Large model size increases operational cost
・High memory consumption in operation

Use Cases

・Advanced analysis of large documents
・Tasks requiring complex logical reasoning
・Development leveraging long-context capabilities

Deep Analysis

Total Parameters

284B

13B activated per token

Context Window

1M tokens

8x increase from V3.2's 128K

Input Price (cache miss)

$0.14/1M tokens

Cache hit: $0.028/1M

Output Price

$0.28/1M tokens

SWE-Bench Verified

79%

vs DeepSeek V4 Pro: 80.6%

GPQA Diamond

88.1%

From official evaluations

Strengths

・Extremely cost-effective with strong coding/agentic performance relative to price
・1M-token context window with efficient hybrid attention architecture (CSA + HCA)
・Open-weight under MIT license with multiple inference providers available

Weaknesses

・Weaker on complex agentic tasks and factual knowledge compared to DeepSeek V4 Pro
・High hallucination rate (96%) when uncertain, as reported by Artificial Analysis
・Preview model: behavior and pricing may change before final release

Competitor Comparison

Model	Arena	SWE	GPQA	Price
DeepSeek V4 Pro	1460	80.6%	90.1%	$0.435/$0.87 (input/output per 1M tokens)
Kimi K2.6	1454	80.2%	90.5%	Not publicly listed
Claude Sonnet 4.6	1459	79.6%	89.9%	Not publicly listed

Sources

Analysis generated: 2026-05-23