TL;DR DeepSeek V4 Pro and V4 Flash are open MoE models with a 1M context window. Pro is built for deep reasoning, coding, and accuracy, while Flash is faster and ~12x cheaper for high-volume tasks. Choose Pro for complex workflows and Flash for speed and scale.

DeepSeek V4 Pro and DeepSeek V4 Flash: What They Are, How They Compare, and How to Use Them

DeepSeek released two new open source language models on April 24, 2026: DeepSeek V4 Pro and DeepSeek V4 Flash. Both use a Mixture of Experts architecture, both support a one million token context window as standard, and both are available through the API and the chat interface at chat.deepseek.com right now.

Below is everything a developer or technical decision maker needs to evaluate them: specifications, pricing, benchmark comparisons against Gemini 3.1 Pro, integration instructions, and answers to the most common questions showing up in search.

DeepSeek V4 Pro vs. DeepSeek V4 Flash: Summary Table

	DeepSeek V4 Pro	DeepSeek V4 Flash
Total Parameters	1.6 trillion	284 billion
Active Parameters	49 billion per token	13 billion per token
Context Window	1 million tokens	1 million tokens
Max Output	384K tokens	384K tokens
Architecture	Mixture of Experts (MoE)	Mixture of Experts (MoE)
Input Price (cache miss)	$1.74 per 1M tokens	$0.14 per 1M tokens
Input Price (cache hit)	$0.145 per 1M tokens	$0.028 per 1M tokens
Output Price	$3.48 per 1M tokens	$0.28 per 1M tokens
Chat Interface Mode	Expert Mode	Instant Mode
API Model Name	`deepseek-v4-pro`	`deepseek-v4-flash`
License	MIT	MIT

What Is DeepSeek V4 Pro?

DeepSeek V4 Pro is the new flagship "frontier tier" model from DeepSeek. It is designed to go head-to-head with the world's most powerful closed-source models, like GPT 5.4 and Gemini 3.1 Pro, while remaining open source under the MIT license.

The Architecture. It uses a massive Mixture of Experts (MoE) design with 1.6 trillion total parameters, but only 49 billion are active per token. This allows it to maintain elite intelligence without the staggering latency or cost of a fully dense 1T+ model.

The Specialty. It is specifically optimized for agentic workflows and complex reasoning. In internal tests, it powers DeepSeek's own coding pipelines, proving its reliability for production-level software engineering.

The Experience. On the chat interface, this is accessed via Expert Mode. It is for when you need the highest possible reasoning ceiling and the most accurate math or code generation.

What Is DeepSeek V4 Flash?

DeepSeek V4 Flash is the high-speed, high-efficiency sibling designed for volume. It is the first "Flash" model to offer a 1M context window at such a low price point.

The Architecture. It features 284 billion total parameters, with just 13 billion active per token. This makes it significantly faster than Pro, with lower latency across the board.

The Specialty. Flash is the workhorse. It is aimed at high-frequency tasks where latency is the priority, like real-time chatbots, basic data extraction, or summarizing long documents.

The Experience. On the chat interface, this is Instant Mode. It is arguably the most capable "small" model on the market, performing on par with the Pro version for standard agentic tasks despite being 12x cheaper to run.

DeepSeek V4 Pro and DeepSeek V4 Flash Benchmarks

The table below shows how both models perform against the current frontier, sourced from DeepSeek's published benchmark comparisons.

Benchmark	DS V4 Pro Max	DS V4 Flash Max	K2.6 Thinking	GLM 5.1 Thinking	Opus 4.6 Max	GPT 5.4 xHigh	Gemini 3.1 Pro High
MMLU Pro	87.5	86.2	87.1	86.0	89.1	87.5	91.0
SimpleQA Verified	57.9	34.1	36.9	38.1	46.2	45.3	75.6
Chinese SimpleQA	84.4	78.9	75.9	75.0	76.2	76.8	85.9
GPQA Diamond	90.1	88.1	90.5	86.2	91.3	93.0	94.3
HLE	37.7	34.8	36.4	34.7	40.0	39.8	44.4
LiveCodeBench	93.5	91.6	89.6	—	88.8	—	91.7
Codeforces (Rating)	3206	3052	—	—	—	3168	3052
HMMT 2026 Feb	95.2	94.8	92.7	89.4	96.2	97.7	94.7
IMOAnswerBench	89.8	88.4	86.0	83.8	75.3	91.4	81.0
Apex	38.3	33.0	24.0	11.5	34.5	54.1	60.9
Apex Shortlist	90.2	85.7	75.5	72.4	85.9	78.1	89.1
MRCR 1M	83.5	78.7	—	—	92.9	—	76.3
CorpusQA 1M	62.0	60.5	—	—	71.7	—	53.8
Terminal Bench 2.0	67.9	56.9	66.7	63.5	65.4	75.1	68.5
SWE Verified	80.6	79.0	80.2	—	80.8	—	80.6
SWE Pro	55.4	52.6	58.6	58.4	57.3	57.7	54.2
SWE Multilingual	76.2	73.3	76.7	73.3	77.5	—	—
BrowseComp	83.4	73.2	83.2	79.3	83.7	82.7	85.9
HLE w/tools	48.2	45.1	54.0	50.4	53.1	52.0	51.6
GDPval AA (Elo)	1554	1395	1482	1535	1619	1674	1314
MCPAtlas Public	73.6	69.0	66.6	71.8	73.8	67.2	69.2
Toolathlon	51.8	47.8	50.0	40.7	47.2	54.6	48.8

DeepSeek V4 Pro leads all current open source models on coding and math benchmarks and is competitive with the best closed source systems. It trails Gemini 3.1 Pro on world knowledge (SimpleQA, MMLU Pro), Claude Opus 4.6 on long context retrieval (MRCR 1M, CorpusQA 1M), and GPT 5.4 on certain agentic tasks (Terminal Bench 2.0, Toolathlon).

DeepSeek V4 Flash tracks within a few points of Pro on most benchmarks, with the largest gaps appearing on agentic tasks and factual retrieval. For standard workloads, the performance difference is unlikely to justify the 12x cost premium of Pro.

How to Use DeepSeek V4 Pro and DeepSeek V4 Flash via the API

Both models use the same base URL as previous DeepSeek models. You do not need to change your endpoint. Just update the model parameter.

Python (OpenAI compatible format):

from openai import OpenAI

client = OpenAI(
    api_key="your-deepseek-api-key",
    base_url="https://api.deepseek.com"
)

# For DeepSeek V4 Pro
response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Your prompt here"}]
)

# For DeepSeek V4 Flash
response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Your prompt here"}]
)

Both models also support the Anthropic API format at https://api.deepseek.com/anthropic.

To enable thinking mode (chain of thought reasoning), pass the thinking parameter in the request body:

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Your prompt here"}],
    extra_body={"thinking": {"type": "enabled"}}
)

Both models support JSON output, tool calls, FIM completion (in non thinking mode), and chat prefix completion.

If you are building agentic applications on top of DeepSeek V4, tools like Thesys let you render model outputs and tool call responses as interactive generative UI components rather than plain text, turning structured data into something users can actually work with.

Explore Generative UI with Thesys

The 1M Context Window: How It Works

Both DeepSeek V4 Pro and DeepSeek V4 Flash ship with a one million token context window as the default. This is not a premium tier or a beta opt in. Every API call and every chat session runs at 1M context.

The technical innovation behind this is a hybrid attention mechanism that DeepSeek calls Compressed Sparse Attention combined with Heavily Compressed Attention. Together with a technique called Manifold Constrained Hyper Connections for residual signal propagation, these reduce the compute and memory cost of long context inference by roughly an order of magnitude on KV cache compared to standard attention.

In practical terms, this means you can send an entire codebase, a full legal contract, or a long conversation history in a single API call without hitting a context ceiling or paying a premium for the extended window.

Migration Deadline: July 24, 2026

The legacy model names deepseek-chat and deepseek-reasoner currently route to DeepSeek V4 Flash in non thinking and thinking modes respectively. But they will be fully retired and inaccessible after July 24, 2026 at 15:59 UTC.

If your production systems reference either of those model names, you need to update them to deepseek-v4-pro or deepseek-v4-flash before that date. This is not a soft deprecation. After the cutoff, those endpoints will stop responding.

Build on DeepSeek V4 with Generative UI

DeepSeek V4 handles the reasoning. Your application still needs an interface. Thesys is a generative UI platform that takes structured model outputs, tool call responses, and agentic workflow results and renders them as interactive, production ready UI components.

Instead of returning raw JSON or plain text to your users, you ship an interface they can see, click, and act on.

If you are building with DeepSeek V4 Pro or DeepSeek V4 Flash, Thesys is worth looking at as the presentation layer.

Try C1 API

FAQs

Is DeepSeek V4 Pro free?

The chat interface at chat.deepseek.com is free to use. You can access DeepSeek V4 Pro through Expert Mode and DeepSeek V4 Flash through Instant Mode at no cost. The API is pay as you go: V4 Pro costs $1.74 per million input tokens (cache miss) and $3.48 per million output tokens. V4 Flash costs $0.14 per million input tokens (cache miss) and $0.28 per million output tokens. DeepSeek also offers a free token grant to new API users.

Can I run DeepSeek V4 Pro locally?

Both models are released as open weights under the MIT license on Hugging Face. V4 Pro is approximately 865 GB, which requires significant hardware to run locally. V4 Flash at around 160 GB is more feasible for self hosting, particularly on high memory machines. Quantized versions from the community are expected to reduce the memory requirements further.

What is the difference between Expert Mode and Instant Mode?

Expert Mode uses DeepSeek V4 Pro. Instant Mode uses DeepSeek V4 Flash. Both are available in the free chat interface at chat.deepseek.com. The distinction is depth versus speed: Expert Mode delivers stronger reasoning at the cost of longer response times, while Instant Mode prioritizes fast, cost efficient responses.

What API model names should I use?

Use deepseek-v4-pro for the flagship model and deepseek-v4-flash for the efficiency model. The old model names deepseek-chat and deepseek-reasoner will stop working on July 24, 2026.

Does DeepSeek V4 support tool calls and function calling?

Yes. Both V4 Pro and V4 Flash support tool calls, JSON output, chat prefix completion, and FIM completion in non thinking mode. They are compatible with the OpenAI ChatCompletions format and the Anthropic API format.

How does DeepSeek V4 Flash compare to V4 Pro?

Flash activates 13 billion parameters per token compared to Pro's 49 billion. It is faster, cheaper, roughly 12x less on output, and performs within a few percentage points of Pro on most benchmarks. DeepSeek says Flash matches Pro on simple agent tasks and closely approaches it on reasoning. For high volume workloads where cost and latency matter more than peak accuracy, Flash is the practical choice.

How does DeepSeek V4 Pro compare to Google Gemini 3.1 Pro?

DeepSeek V4 Pro leads on coding and math, with 93.5 on LiveCodeBench, 3206 on Codeforces, and 89.8 on IMOAnswerBench. Gemini 3.1 Pro leads on world knowledge and general reasoning, with 75.6 on SimpleQA Verified, 91.0 on MMLU Pro, and 44.4 on HLE. For coding and agentic workloads, V4 Pro is competitive or ahead. For broad factual knowledge, Gemini remains stronger. V4 Pro is also considerably cheaper at $3.48 per million output tokens.

Related articles

May 2026 Security Incident

Why We're Open Sourcing OpenUI

AI Copilots: A Complete Guide to Intelligent Assistance

What Are AI Agents: Types, Capabilities, and Real-World Use

ChatKit: What It Is, What It Isn’t, and Where C1 Fits In

PowerPoint Automation Reimagined with C1 Artifacts API

The Future of AI Interfaces is Generative

MCP UI: Making AI Apps Interactive

What Is AG-UI and Why It Matters for AI-Native User Interfaces

How to design AI-Native Conversational Interfaces : From Templates to Generative UI

GPT 5 vs. GPT 4.1

How to Build Generative UI Applications

Implementing Generative Analytics with Thesys and MCP

Evolution of Analytics: From Static Dashboards to Generative UI

Why Generating Code for Generative UI is a bad idea

Building the First Generative UI API: Technical Architecture and Design Decisions Behind C1

How we evaluate LLMs for Generative UI

Generative UI vs Prompt to UI vs Prompt to Design

Generative UI: What It Is and Why It Matters