How do you try DeepSeek-V4?

DeepSeek released a preview of its V4 models on April 24. There are two: DeepSeek-V4-Pro, a large model with 1.6 trillion parameters in total, and DeepSeek-V4-Flash, a smaller, faster and cheaper one with 284 billion. Both are mixture-of-experts models, meaning only a fraction of the model (49 billion and 13 billion parameters respectively) works on each piece of text, which keeps them cheaper to run than their size suggests. Both accept up to one million tokens of context (a token is a chunk of text, roughly three-quarters of a word), and both have their weights published openly.

"Preview" means DeepSeek is still finishing the models and may change them. Treat results as a first look, not a final verdict.

Three ways to try it

1. In the chat app. Go to chat.deepseek.com or open the DeepSeek mobile app. DeepSeek says V4 is available there through two modes, Expert Mode and Instant Mode; try the same question in both and compare. Start with a task you know well, so you can judge the answer.

2. Through the API. If you already use DeepSeek's API, the address stays the same; only the model name changes, to deepseek-v4-pro or deepseek-v4-flash. The API accepts requests in OpenAI's Chat Completions format and in Anthropic's format, so many existing tools work after a settings change. The older names deepseek-chat and deepseek-reasoner now point to V4-Flash and are scheduled to stop working after July 24, 2026, so update any scripts that use them.

3. On your own terms. The weights are on Hugging Face, so other companies can host the models. These are far too large for a laptop, but cloud providers and model platforms can run them, under their own data policies and data-center locations.

A worked example: comparing the two models with thinking on

Both models can think before answering: they write out a chain of reasoning, then the final reply. According to DeepSeek's guide, thinking is on by default, effort can be set to high or max, and with the OpenAI Python library the switch goes in extra_body. This script asks both models the same question and prints the reasoning length and the answer.

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["DEEPSEEK_API_KEY"],
                base_url="https://api.deepseek.com")

QUESTION = ("A club has 48 members. 3/8 come on Tuesdays, and a third of "
            "those also come on Thursdays. How many come on both days?")

for model in ["deepseek-v4-flash", "deepseek-v4-pro"]:
    reply = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": QUESTION}],
        reasoning_effort="high",
        extra_body={"thinking": {"type": "enabled"}},
    )
    msg = reply.choices[0].message
    print(model, "| reasoning chars:", len(msg.reasoning_content or ""))
    print(msg.content, "\n")

The correct answer is 6: three-eighths of 48 is 18, and a third of 18 is 6. Run it with your own test questions too, taken from your real work, and note which model gets them right, how long each takes, and how much each costs on DeepSeek's pricing page. Two details from the guide matter if you build further: in thinking mode, settings such as temperature are ignored, and if the model calls tools, you must send its reasoning back in later turns or the request fails.

What should you keep out of it?

DeepSeek's privacy policy says the company collects what you type and upload, your chat history, and technical details such as your device, IP address and even typing patterns. It lists improving its technology among the uses of that data, says the information is stored on servers in the People's Republic of China, and allows sharing with authorities when the company judges it necessary. Read the current version of the policy yourself before you rely on the service.

In practice, that means keeping these out of the hosted app and API:

• Personal information about you or others: health, finances, addresses, ID numbers.
• Client or customer data, and anything covered by a confidentiality agreement.
• Passwords, API keys and internal documents.
• Information your employer or local privacy law says must stay in a particular country.

Questions about public topics, practice writing, puzzles and code with nothing secret in it are reasonable things to test. If you want to use V4 with sensitive material, run the open weights through a provider whose data location and terms you have checked, or on infrastructure you control. You can also delete your chat history in DeepSeek's settings, but deleting it later is not the same as never sharing it.