Updated
DeepSeek
DeepSeek V4 Pro
DeepSeek's largest V4 model for coding, knowledge and STEM; Super Agent runs the April preview.
What is DeepSeek V4 Pro?
DeepSeek V4 Pro is the large DeepSeek V4 model, released as an open-weight preview on April 24, 2026, with 1.6T total parameters and 49B active per token.
DeepSeek built it for agentic coding, broad world knowledge and reasoning in math, STEM and code, and at launch called its maximum-effort mode the best open-source model available. Its hybrid attention needs 27% of the per-token inference compute and 10% of the KV cache of DeepSeek V3.2 at 1M tokens of context. It accepts text only. Super Agent runs the April preview build through third-party hosts; DeepSeek has since released V4 Pro 0813, which its own API now serves under the same name.
- Provider
- DeepSeek
- Context window
- 1M tokens
- Input
- Text
- Reasoning
- Off, medium, high, extra high
- Released
- Apr 24, 2026
How DeepSeek V4 Pro performs
Published by DeepSeek
Results DeepSeek reported for DeepSeek V4 Pro. Each lab runs its own setup, so compare these with other models only inside the same table.
| Benchmark | Score | Setting |
|---|---|---|
| GPQA Diamond | 89.1% | High thinking, April preview |
| Humanity's Last Exam | 34.5% | High thinking, April preview |
| LiveCodeBench | 89.8% | High thinking, April preview |
| Codeforces | 2,919 rating | High thinking, April preview |
| HMMT 2026 Feb | 94.0% | High thinking, April preview |
| SWE-bench Verified | 79.4% | High thinking, April preview |
| SWE-bench Pro | 54.4% | High thinking, April preview |
| Terminal-Bench 2.0 | 63.3% | High thinking, April preview |
| MRCR 1M | 83.3% | High thinking, April preview |
| BrowseComp | 80.4% | 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
#21 of 27 in Super Agent
All text prompts, with style control on. #64 of 418 on the full leaderboard, from 57,588 votes.
- 1Claude Opus 5.51509
- 2Claude Fable 51503
- 3Claude Fable 5.11503
- 4Claude Opus 51491
- 5Gemini 3.7 Flash1490
- 21DeepSeek V4 Pro1458
Coding
#25 of 27 in Super Agent
Text prompts that ask for code. #79 of 413 on the full leaderboard, from 17,142 votes.
- 1Claude Fable 51551
- 2Claude Opus 5.51548
- 3GPT-6 Astra1543
- 4Kimi K31539
- 5Claude Fable 5.11538
- 25DeepSeek V4 Pro1501
WebDev
#25 of 27 in Super Agent
Building working web apps in Code Arena. #76 of 149 on the full leaderboard, from 13,974 votes.
- 1Claude Opus 5.51807
- 2GPT-6 Astra1775
- 3Claude Sonnet 5.51768
- 4GPT-6.1 Sol1749
- 5Claude Fable 5.11743
- 25DeepSeek V4 Pro1447
Math
#24 of 27 in Super Agent
Text prompts with math problems. #87 of 404 on the full leaderboard, from 3,066 votes.
- 1Claude Fable 5.11529
- 2Claude Fable 51521
- 3Claude Opus 51517
- 4GPT-6.1 Sol1513
- 5Claude Opus 5.51512
- 24DeepSeek V4 Pro1444
Hard prompts
#22 of 27 in Super Agent
Complex, multi-part and demanding text prompts. #65 of 418 on the full leaderboard, from 38,511 votes.
- 1Claude Opus 5.51535
- 2Claude Fable 51529
- 3Claude Fable 5.11523
- 4Kimi K31516
- 5Claude Opus 51515
- 22DeepSeek V4 Pro1480
Creative writing
#17 of 27 in Super Agent
Stories, poems and other original writing. #45 of 416 on the full leaderboard, from 9,731 votes.
- 1Claude Opus 5.51520
- 2Claude Fable 51501
- 3Gemini 3.7 Flash1492
- 4Claude Fable 5.11487
- 5Gemini 3.1 Pro Preview1481
- 17DeepSeek V4 Pro1446
Source: LMArena, leaderboard published . Scores move as new votes come in.
What you can do with DeepSeek V4 Pro
1.6T total parameters, 49B active per token
Max-effort mode named the best open-source model at launch, per DeepSeek
Leads open models on world knowledge, trailing only Gemini 3.1 Pro, per DeepSeek
Hybrid attention keeps its 1M-token context efficient
How to prompt DeepSeek V4 Pro
- Paste long source text and ask it to pull out what matters.
- Turn reasoning on for hard tasks; Super Agent starts it at off.
- Describe visual references in words, since it reads text only.
Where DeepSeek V4 Pro fits best
Synthesizing large amounts of text
Long multi-step automation
Hard reasoning in math and code
Analysis of long documents
How DeepSeek V4 Pro compares
| Model | LMArena text | LMArena coding | Context window | Reasoning |
|---|---|---|---|---|
| DeepSeek V4 Pro | 1458#64 | 1501#79 | 1M tokens | Off, medium, high, extra high |
| GPT-6 Sol | 1459#62 | 1519#38 | 1.05M tokens | Off to max |
| Grok 4.6 | 1456#66 | 1510#60 | 500K tokens | Low to extra high |
| Gemini 3.5 Flash Lite | 1455#71 | 1502#77 | 1M tokens | Low, medium, high |
| Kimi K2.6 | 1461#59 | 1516#44 | 256K tokens | On or off |
Good to know before you start
- Super Agent runs the April 2026 preview, not DeepSeek's newer GA build (V4-Pro-0813)
- Can't read images; the model picker marks it Text only
- Reasoning jumps from off straight to medium, with no low setting
Start chatting with DeepSeek V4 Pro
Chat with DeepSeek V4 Pro in Super AgentSelect DeepSeek V4 Pro
Open Super Agent on Flashloop and choose DeepSeek V4 Pro 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 Pro answers in Super Agent and runs any generations you asked for.
Questions about DeepSeek V4 Pro
Which DeepSeek V4 Pro build does Super Agent use?
The April 2026 preview. DeepSeek released the V4 Pro GA build (V4-Pro-0813) on August 13, 2026, and its own API now serves it as deepseek-v4-pro. The preview reaches Super Agent through third-party hosts.
How is DeepSeek V4 Pro different from V4 Flash?
Both share the same architecture. V4 Pro is much larger (1.6T total, 49B active parameters versus 284B and 13B) and aimed at harder reasoning, while V4 Flash is built for speed.
Can DeepSeek V4 Pro read images?
No. V4 Pro reads text only, and the model picker shows a Text only tag for it.
Which reasoning levels does DeepSeek V4 Pro support?
The picker offers off, medium, high and extra high and starts at off. DeepSeek's card for the April preview lists non-thinking, high and max modes. DeepSeek's own API, which now serves the newer build, maps medium and extra high to high.
More chat models on Flashloop
GPT-6 SolOpenAI's GPT-6 model between Luna and Astra, trained like Astra and able to reply with reasoning off.
Grok 4.6xAI's Grok 4.6 for turning a broad product idea into a working first version, with 500K context.
Gemini 3.5 Flash LiteGoogle's fastest, lowest-cost Gemini 3.5 model for extraction, routing and subagent tasks.
Kimi K2.6Moonshot AI's open coding model that turns prompts and visual inputs into working interfaces.