How to Run DeepSeek V4 Pro: Every Working Option

DeepSeek V4 Pro is the big sibling in the V4 family: 1.6 trillion total parameters, 49B active per token, 1M context, and open MIT weights like everything DeepSeek ships. With today's Flash 0731 release stealing the headlines, here is the matching guide for Pro: every way to use it, what it costs, and the straight answer on running it yourself.

Your Options at a Glance

Option 1: The Official API

Same OpenAI compatible surface as every DeepSeek model; only the model string differs:

from openai import OpenAI

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

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Prove or refute this invariant, step by step."}],
)
print(resp.choices[0].message.content)

Pricing: $0.435 per million input on cache miss, $0.003625 on cache hits, $0.87 per million output. Context is 1M tokens in, up to 384K out, thinking mode on by default. Notes that matter: the API still serves the Preview build (today's 0731 update touched Flash only), and the Responses API support that Flash 0731 got has not come to Pro. If you are wiring up Codex style tooling, that is a point for Flash.

Option 2: LU Labs (flat price)

If you want V4 Pro without an API console, LU Labs carries it in the model picker on the Hosted Pro plan (49€/mo), right next to V4 Flash, Qwen, GLM and the rest of the open lineup. Flat monthly quota instead of per token billing, EU based, privacy first, and it plugs into the same studio experience as the local app. For most people this is the cheapest way to poke a 1.6T model seriously: an evening of heavy V4 Pro output on the raw API adds up faster than you expect at $0.87 per million.

Option 3: OpenRouter

Listed as deepseek/deepseek-v4-pro. One string change if you are already there; unified billing; a routing margin on top. With MIT weights public, independent hosts can and do appear under the listing, which over time pressures prices in a way the closed flagships never see.

Option 4: Self Hosting (the straight answer)

The weights are on Hugging Face under MIT, so nothing legally stops you. Physically is another story. At 4 bit, 1.6T parameters approach 900 GB of weights before KV cache; the recommended serving setups are multi GPU nodes (think H200 or B200 class, several of them, with expert parallelism). This is server room math: no Mac Studio, no workstation, no dual GPU gaming rig gets there. It is the same conclusion as our Kimi K3 hardware post, one weight class lower.

The family answer is division of labor: V4 Flash is the one you run at home (155 GB at 4 bit, 13B active, and after 0731 it beats Pro Preview on several agent benchmarks anyway), V4 Pro is the one you rent for the hardest reasoning. The full head to head is in Flash vs Pro, and the Flash local walkthrough is here.

Which Option for Whom