DeepSeek V4 Flash in the Cloud: LU Labs, API, OpenRouter

The official DeepSeek-V4-Flash-0731 build went live today, and not everyone wants to download 155 GB of weights to try it (that path is covered in the local guide). Here is every cloud route that works right now, with real prices and working code.

Your Options at a Glance

Option 1: LU Labs (one credit balance, no key juggling)

LU Labs is the hosted layer from the team behind Locally Uncensored: the same studio experience (chat, code, image, video), with the heavy models running on rented GPUs instead of yours. The model picker includes the whole DeepSeek line, and V3.2, V4 Flash and V4 Pro are all in it on every plan and on a credit pack from 15€, metered in credits against one balance. A few reasons people pick this over raw API keys:

• One balance you can see. A plan comes with monthly credits and a pack with a fixed amount, and a long agent loop draws on that balance.
• Open weights, EU billing. Served for the plan price, billed in the EU, and no training on your prompts.
• One picker, whole lineup. V4 Flash next to Qwen, GLM, Kimi and the unfiltered community builds, switchable per conversation.
• Works with the desktop app. Run small models locally in Locally Uncensored and borrow cloud muscle for the big ones, same interface.

Option 2: The Official DeepSeek API

The direct route. OpenAI compatible, so any client works with a changed base URL, key, and model name:

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-flash",
    messages=[{"role": "user", "content": "Refactor this function and explain the change."}],
)
print(resp.choices[0].message.content)

Pricing as of today: $0.14 per million input tokens (cache miss), a remarkable $0.0028 per million on cache hits, and $0.28 per million output. Both V4 models take 1M tokens of input context and can emit up to 384K output tokens. Two fine print items worth knowing: DeepSeek has announced 2x pricing during Beijing peak hours (not yet in effect, date pending), and thinking mode is the default (the model reasons before answering; you can switch it off per request).

The 0731 build also natively supports the Responses API format and is adapted for Codex, which is a first for DeepSeek: agent tools that speak OpenAI's Responses interface run against deepseek-v4-flash without a compatibility shim. Given the agent benchmarks this build posts (82.7 on Terminal Bench 2.1), that pairing is the headline feature. Details in the 0731 explainer.

Option 3: OpenRouter

OpenRouter lists deepseek/deepseek-v4-flash, so apps already wired to it switch with one model string. You get unified billing and fallbacks; you pay the routing margin, and with open weights now published, expect independent hosts to appear under the same listing and push prices around. For a single model, the official API or a plan with credits is usually the simpler deal.

Which Cloud Route for Whom

And V4 Pro?

Everything here applies to the big sibling too, at higher prices ($0.435 / $0.87 per million on the official API) and with no realistic local escape hatch: V4 Pro is a 1.6T parameter MoE. We wrote up its options in How to Run DeepSeek V4 Pro and the head to head in V4 Flash vs V4 Pro.

In the handbook

• All DeepSeek guides, every generation on one page
• Install and update, the SmartScreen warning and how updates arrive
• Chat, the model picker, the context window and thinking mode
• FAQ and glossary, answers to the questions people really ask