Updated
DeepSeek
DeepSeek V4 Flash
DeepSeek's fast, economical V4 model; Super Agent runs the April 2026 preview build.
What is DeepSeek V4 Flash?
DeepSeek V4 Flash is the smaller DeepSeek V4 model, released as an open-weight preview on April 24, 2026, with 284B total parameters and 13B active per token.
DeepSeek built it for fast, economical responses with reasoning close to V4 Pro, and its new attention design keeps the 1M-token context cheap to run. It accepts text only. Super Agent runs that April preview build through third-party hosts. DeepSeek has since released newer builds, including V4 Flash 0731 and V4.1 Flash, and the official results on this page are for the preview.
- Provider
- DeepSeek
- Context window
- 1M tokens
- Input
- Text
- Reasoning
- Off, medium, high, extra high
- Released
- Apr 24, 2026
How DeepSeek V4 Flash performs
Published by DeepSeek
Results DeepSeek reported for DeepSeek V4 Flash. Each lab runs its own setup, so compare these with other models only inside the same table.
| Benchmark | Score | Setting |
|---|---|---|
| GPQA Diamond | 87.4% | High thinking, April preview |
| Humanity's Last Exam | 29.4% | High thinking, April preview |
| LiveCodeBench | 88.4% | High thinking, April preview |
| Codeforces | 2,816 rating | High thinking, April preview |
| HMMT 2026 Feb | 91.9% | High thinking, April preview |
| SWE-bench Verified | 78.6% | High thinking, April preview |
| SWE-bench Pro | 52.3% | High thinking, April preview |
| Terminal-Bench 2.0 | 56.6% | High thinking, April preview |
| MRCR 1M | 76.9% | High thinking, April preview |
| BrowseComp | 53.5% | High thinking, April preview |
Source: huggingface.co
On LMArena
People compare two anonymous answers and vote for the better one. Ranked here against the other chat models you can pick in Super Agent.
Text
#27 of 27 in Super Agent
All text prompts, with style control on. #105 of 418 on the full leaderboard, from 51,982 votes.
- 1Claude Opus 5.51509
- 2Claude Fable 51503
- 3Claude Fable 5.11503
- 4Claude Opus 51491
- 5Gemini 3.7 Flash1490
- 27DeepSeek V4 Flash1436
Coding
#27 of 27 in Super Agent
Text prompts that ask for code. #106 of 413 on the full leaderboard, from 15,295 votes.
- 1Claude Fable 51551
- 2Claude Opus 5.51548
- 3GPT-6 Astra1543
- 4Kimi K31539
- 5Claude Fable 5.11538
- 27DeepSeek V4 Flash1484
WebDev
#15 of 27 in Super Agent
Building working web apps in Code Arena. #37 of 149 on the full leaderboard, from 4,948 votes.
- 1Claude Opus 5.51807
- 2GPT-6 Astra1775
- 3Claude Sonnet 5.51768
- 4GPT-6.1 Sol1749
- 5Claude Fable 5.11743
- 15DeepSeek V4 Flash1582
Math
#27 of 27 in Super Agent
Text prompts with math problems. #124 of 404 on the full leaderboard, from 2,702 votes.
- 1Claude Fable 5.11529
- 2Claude Fable 51521
- 3Claude Opus 51517
- 4GPT-6.1 Sol1513
- 5Claude Opus 5.51512
- 27DeepSeek V4 Flash1424
Hard prompts
#27 of 27 in Super Agent
Complex, multi-part and demanding text prompts. #102 of 418 on the full leaderboard, from 34,766 votes.
- 1Claude Opus 5.51535
- 2Claude Fable 51529
- 3Claude Fable 5.11523
- 4Kimi K31516
- 5Claude Opus 51515
- 27DeepSeek V4 Flash1459
Creative writing
#27 of 27 in Super Agent
Stories, poems and other original writing. #96 of 416 on the full leaderboard, from 8,579 votes.
- 1Claude Opus 5.51520
- 2Claude Fable 51501
- 3Gemini 3.7 Flash1492
- 4Claude Fable 5.11487
- 5Gemini 3.1 Pro Preview1481
- 27DeepSeek V4 Flash1408
Source: LMArena, leaderboard published . Scores move as new votes come in.
What you can do with DeepSeek V4 Flash
284B total parameters, 13B active per token
Reasoning close to V4 Pro, according to DeepSeek
New attention design for cheaper 1M-token context
Open weights under the MIT license
How to prompt DeepSeek V4 Flash
- Use it for many quick text tasks in a row, where speed matters most.
- Describe images in words, since it can't read attachments that aren't text.
- Turn reasoning on for anything that needs a plan; it starts at off.
Where DeepSeek V4 Flash fits best
High-volume prompt writing
Quick text chat
Summaries of long text
Fast drafts you will refine later
How DeepSeek V4 Flash compares
| Model | LMArena text | LMArena coding | Context window | Reasoning |
|---|---|---|---|---|
| DeepSeek V4 Flash | 1436#105 | 1484#106 | 1M tokens | Off, medium, high, extra high |
| Grok 4.7 | 1441#96 | 1493#91 | 500K tokens | Low to max |
| GPT-6 Luna | 1442#91 | 1507#65 | 1.05M tokens | Off to max |
| GPT 5.6 Luna | 1453#75 | 1502#76 | 1.05M tokens | Off to max |
| Gemini 3.5 Flash Lite | 1455#71 | 1502#77 | 1M tokens | Low, medium, high |
Good to know before you start
- Super Agent runs the April 2026 preview; newer DeepSeek builds exist
- Text input only; the model picker tags it Text only
- DeepSeek says it trails V4 Pro slightly on pure knowledge and the hardest agent workflows
Start chatting with DeepSeek V4 Flash
Chat with DeepSeek V4 Flash in Super AgentSelect DeepSeek V4 Flash
Open Super Agent on Flashloop and choose DeepSeek V4 Flash in the model picker.
Enter your prompt
Ask for what you need: an answer, a plan, or the images and videos you want the agent to make.
Send and keep going
Send the message. DeepSeek V4 Flash answers in Super Agent and runs any generations you asked for.
Questions about DeepSeek V4 Flash
Which DeepSeek V4 Flash build does Super Agent use?
The April 2026 preview, which Super Agent reaches through third-party hosts. DeepSeek has since shipped V4 Flash 0731, V4.1 Flash and V4.1 Flash 1017. Its own API now serves V4.1 Flash as deepseek-flash and, for now, routes the old deepseek-v4-flash name there too.
What is DeepSeek V4 Flash built for?
Fast, economical text work. DeepSeek says its reasoning comes close to V4 Pro and that it performs comparably on simple agent tasks.
Can DeepSeek V4 Flash read images?
No. It accepts text only, and Super Agent tags it Text only in the model picker.
Which reasoning levels does DeepSeek V4 Flash support?
Super Agent lists off, medium, high and extra high, starting at off. DeepSeek documents the April preview with three modes: non-thinking, high and max. Its current API offers low, high and max and maps medium and extra high to high.
More chat models on Flashloop
Grok 4.7xAI's newest Grok, trained on many-hour problems and built to check its own work.
GPT-6 LunaOpenAI's most efficient GPT-6 model for focused, high-volume work, and Super Agent's default.
GPT 5.6 LunaOpenAI's fastest, most affordable GPT-5.6 model for quick chat and high-volume work.
Gemini 3.5 Flash LiteGoogle's fastest, lowest-cost Gemini 3.5 model for extraction, routing and subagent tasks.