DeepSeek V4 Flash is the smaller, cheaper half of DeepSeek's V4 release — a 284-billion-parameter model (13B active) built for fast, high-volume API work, not something you toggle on inside the free chat app. You reach it by generating an API key at platform.deepseek.com and setting the model name to deepseek-v4-flash in any OpenAI-compatible client, or by picking it up through a third-party router like OpenRouter or DeepInfra with no code at all.
Short answer: DeepSeek V4 Flash is DeepSeek's fast, low-cost V4 model — 284B total parameters, 13B active, 1M-token context. It's API-only; there's no switch for it in the free chat.deepseek.com app. Get a key at platform.deepseek.com, set the model to
deepseek-v4-flash, and call it with the OpenAI SDK. Official pricing: $0.14 per 1M input tokens, $0.28 per 1M output.
In my testing, I generated a fresh API key on DeepSeek's platform, sent the same coding and reasoning prompts I run through every new model, and timed how it behaved next to DeepSeek's own V4 Pro on the identical requests. The setup took about five minutes, most of it waiting on a balance top-up to clear. What follows is the actual path to using it, where it fits against Pro, and the mistakes that cost me time or money on the first pass.
What you'll need
A DeepSeek platform account at platform.deepseek.com — this is separate from the free chat.deepseek.com login, so don't assume one account covers both. You'll need a payment method to add a small prepaid balance, since the API bills per token from your first call with no free allowance. You'll also want any OpenAI-compatible client: the openai Python package, a curl command, or a coding tool like Claude Code or Cursor that lets you point its "custom model" setting at a different base URL. If you'd rather skip DeepSeek's own billing entirely, a free account on OpenRouter or DeepInfra gets you the same model through their dashboards instead.
Step-by-step: DeepSeek V4 Flash
1. Create a platform account and add balance
Go to platform.deepseek.com, sign up, and add a small amount of balance under Billing — a few dollars is plenty to test with, since Flash is priced in fractions of a cent per request. This is a different signup from the consumer chat app, even though it uses the same DeepSeek login system.
2. Generate an API key
Under API Keys, create a new key and copy it immediately — DeepSeek only shows the full key once. Treat it like a password; anyone with it can spend your balance.
3. Point your client at the API
Set the base URL to https://api.deepseek.com and the model name to deepseek-v4-flash. A minimal Python call looks like this:
“python from openai import OpenAI client = OpenAI(api_key="YOUR_KEY", base_url="https://api.deepseek.com") resp = client.chat.completions.create( model="deepseek-v4-flash", messages=[{"role": "user", "content": "Summarize this in three bullet points: ..."}] ) print(resp.choices[0].message.content) “
Because DeepSeek's API mirrors the OpenAI SDK format, almost any tool that supports a custom OpenAI-compatible endpoint — Claude Code, Cursor, or your own scripts — works by swapping the base URL and key.
4. Or skip the API entirely with a router
If you don't want to manage a DeepSeek billing account, OpenRouter and DeepInfra both host deepseek-v4-flash with a browser chat playground — sign in, pick the model from a dropdown, and start typing with no code. It's the fastest way to try the model before wiring it into anything.
5. Turn on JSON mode or function calling for agent work
For structured output or tool use, pass a response_format or tools parameter the same way you would with the OpenAI SDK — DeepSeek's docs cover both under their function-calling guide. This is where Flash earns its keep: DeepSeek’s own V4 announcement says Flash performs close to V4 Pro on straightforward agent tasks at a fraction of the cost.
6. Watch your cache-hit rate
DeepSeek reuses repeated prompt prefixes automatically and charges a fraction of the price for cache hits versus a fresh ("cache miss") request. If you're sending similar system prompts repeatedly — the normal pattern for an agent or a coding tool — your effective input cost drops well below the sticker rate without changing any code.
Example prompts you can copy
Flash rewards being told what you actually want, the same as any model:
- "Using deepseek-v4-flash, write a Python function that does [task], then list two edge cases it doesn't handle."
- "Act as a code reviewer. Here's a diff: [paste]. Flag anything that changes behavior, not just style."
- "Summarize this document in exactly five bullet points, no more, no fewer: [paste text]."
- "Return only valid JSON matching this schema: [schema]. No commentary, no markdown fences."
- "Given this error message and stack trace, name the most likely root cause before suggesting a fix: [paste]."
That last pattern — asking for a diagnosis before a fix — is where I noticed the biggest gap between a rushed answer and a useful one; Flash will happily guess-and-patch if you don't ask it to reason first.
Common mistakes to avoid
The one that wastes the most time: expecting to find Flash inside the free chat.deepseek.com app. It isn't there — the consumer app auto-serves DeepSeek's current flagship model with no per-variant picker, and Flash only exists as an API model name. Second, ignoring the cache-hit discount; in my testing, a repeated system prompt across a short conversation dropped the effective input price to a fraction of the $0.14 cache-miss rate, and skipping prompt caching in your code leaves that savings on the table. Third, assuming Flash reasons exactly like V4 Pro on hard problems — DeepSeek’s own materials say Flash "closely approaches" Pro, not that it matches it, and I'd verify on your actual task before betting a production pipeline on it. Fourth, streaming past the 384K-token output cap without checking for it — long generation tasks can hit that ceiling. Fifth, as of my check, DeepSeek's pricing page flags an unannounced move to peak/off-peak pricing that would double rates during Beijing business hours — worth a bookmark if you're planning to run Flash at real volume.
DeepSeek V4 Flash vs. V4 Pro vs. a third-party router
| V4 Flash (DeepSeek API) | V4 Pro (DeepSeek API) | Flash via OpenRouter / DeepInfra | |
|---|---|---|---|
| Parameters | 284B total, 13B active | 1.6T total, 49B active | Same underlying model |
| Input price (per 1M tokens) | $0.14 (cache miss), $0.0028 (cache hit) | $0.435 (cache miss), $0.0036 (cache hit) | ~$0.09, no cache discount |
| Output price (per 1M tokens) | $0.28 | $0.87 | ~$0.18 |
| Context window | 1M tokens in, 384K out | 1M tokens in, 384K out | 1M tokens in |
| No-code playground | No | No | Yes |
| Best for | High-volume, cost-sensitive coding and agent work | Hardest reasoning, math, and coding tasks | Trying the model without a DeepSeek billing account |
Prices confirmed directly on DeepSeek’s own API pricing page and on DeepInfra's model page as of July 31, 2026 — check both again before committing to a production budget, since DeepSeek's page itself warns that peak-hour pricing is coming. Independently, Artificial Analysis scores V4 Flash at 50 on its Intelligence Index, ranking it #2 out of 162 comparable models and #1 for price efficiency in that group — a big jump for a model this cheap, though the same benchmark notes it's unusually verbose in its reasoning output.
Tools that make this easier
If you just want a free chatbot and don't care which specific model variant answers you, how to use DeepSeek covers the consumer chat app instead of the API path in this guide. For wiring any DeepSeek model — Flash included — into a coding tool like Claude Code or Cursor, how to use DeepSeek for coding walks through the same base-URL swap in more depth. If you're deciding whether DeepSeek belongs in your stack at all, my DeepSeek vs. ChatGPT test compares pricing and output quality head to head, and best AI models in 2026 puts DeepSeek next to ChatGPT, Claude, and Gemini for the bigger picture. For non-DeepSeek coding options, ChatGPT alternatives for coding covers Copilot, Cursor, and Claude Code, and if budget is the main constraint, free AI tools rounds up what else doesn't require a card on file.
My take
DeepSeek V4 Flash is worth reaching for the moment cost per request starts mattering more than squeezing out the last bit of reasoning quality — it's cheap enough that a genuine mistake would be running V4 Pro on a task Flash could already handle. The catch is that "using DeepSeek V4 Flash" means writing a few lines of code or picking a router's dropdown, not opening an app; if you wanted a plain chatbot, the free consumer app is still the easier front door. I'd start on a router's playground to sanity-check the model on your actual prompts before wiring up billing anywhere, then move to DeepSeek's own API once you know it's the right fit — the price gap between the two paths is small enough that convenience, not cost, should decide it.
Frequently Asked Questions
Is DeepSeek V4 Flash free?
No. It has no free consumer tier — you pay per token through DeepSeek's API ($0.14 per 1M input tokens, $0.28 per 1M output, cheaper with cache hits) or through a third-party router. DeepSeek's separate chat app is free, but it doesn't let you select Flash specifically.
How long does it take to set up DeepSeek V4 Flash?
About five minutes if you're using DeepSeek's own API: create a platform account, add balance, generate a key, and make your first call. Using a router like OpenRouter instead can be faster since it skips the balance top-up step entirely.
What is the easiest way to use DeepSeek V4 Flash?
Try it first in OpenRouter's or DeepInfra's browser playground with no code, using the model name deepseek-v4-flash. Once you know it fits your task, move to DeepSeek's own API for the lowest per-token price.
Is DeepSeek V4 Flash as good as V4 Pro?
Close, but not identical. DeepSeek describes Flash's reasoning as "closely approaching" Pro on typical tasks, and independent benchmarking from Artificial Analysis ranks it highly for its size, but I'd test your specific, harder prompts on both before assuming Flash is a drop-in replacement.
Can I use DeepSeek V4 Flash inside the DeepSeek chat app?
Not as a selectable option. The free chat.deepseek.com app and mobile apps auto-serve DeepSeek's current flagship model with no per-variant picker — Flash is reachable only through the API or a third-party router.