Free DeepSeek API access: the honest map.

Quick answer: what is and is not free.

The honest framing for this family is different from most "free API" searches: DeepSeek's paid rates are so low that free tiers barely undercut them. A single dollar covers roughly 2,500 typical chat turns on V4 Flash at published rates, which means the real question is rarely how to avoid paying, it is how to start without ceremony. That is what a no-card starting credit is for.

• DeepSeek's own platform: no free tier. Access is pay-as-you-go per token from the first request. The consolation is that the rates are the lowest of any capable family.
• APIsRouter starting credit: free credit, no card, spendable on deepseek-v4-flash and deepseek-v4-pro at catalog rates. After the credit, pay-as-you-go top-ups; an idle month costs nothing.
• Third-party free variants: aggregator platforms periodically list DeepSeek-based models with a :free tag, rate-limited and rotating. Fine for sampling, fragile to build on.
• Self-hosting: V4's weights are MIT-licensed and public. V4 Flash at 284B total parameters is the smaller variant, and it still wants multi-GPU server hardware.

What DeepSeek offers officially: no free tier, very low rates.

DeepSeek's API docs publish per-token prices with no free allowance attached: V4 Flash at $0.14 per million input tokens and $0.28 output, V4 Pro at $0.435 and $0.87 as of July 2026, with automatic context caching discounting repeated input to $0.0028 and $0.003625 per million on cache hits. There is no trial quota or free daily cap on the platform. That pricing posture is deliberate: the company competes by making metered access nearly free-shaped rather than by gating a free tier. For perspective, the entire monthly bill of a busy hobby assistant, ten million input tokens and two million out, lands around two dollars on Flash at published rates. Most projects hunting for a free DeepSeek API would spend more engineering time working around a free tier's caps than that.

The starting-credit route: free to start, metered after, no card.

APIsRouter's free route works like this: a starting credit issued without a card, spendable on any catalog id including both DeepSeek V4 variants, at the same catalog rates as paid traffic. No degraded free endpoint, no separate model list, no daily reset to schedule around. When the credit is spent, access continues as pay-as-you-go top-ups with no subscription. To be exact about the boundary: this is not unlimited free DeepSeek, because unlimited free inference does not exist anywhere GPUs cost money. What the credit covers comfortably is evaluation and early development: at catalog Flash rates, even a small credit funds thousands of chat turns, enough to test prompts, measure real token usage in the console, and know your unit costs before the first top-up. It also covers the frontends this search often comes from: the endpoint is a standard OpenAI-compatible /v1, so SillyTavern, RisuAI, and similar custom-endpoint apps run on the credit the same way scripts do. For a feel of what evaluation traffic costs against a credit, work one evening of roleplay at DeepSeek's published Flash rates: forty turns, each resending 4,000 tokens of context and returning a 300-token reply, is 160K input and 12K output tokens, roughly 2.6 cents. A week of prompt-engineering iterations on a support-bot flow lands in the same range. Evaluation genuinely is the free part of this family's economics; the credit exists so the meter never blocks it.

curl https://api.apisrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $APISROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Draft a polite follow-up email for an unanswered invoice."}]
  }'

The rotating-free-variant route: fine for sampling.

Aggregator platforms periodically list DeepSeek-based models under free tags, rate-limited per minute and per day on a shared account. Those listings are real, and for sampling a model before any commitment they work. Their limits are also structural: the set of free variants rotates without notice, daily request caps land mid-session by design, and free-variant traffic may route to hosts with prompt-logging terms disclosed per provider. Treat the route accordingly: good for a first impression, wrong for anything with state, users, or a schedule. A hardcoded free-variant id is a dependency on a promotion. If a project outgrows sampling, the exit is the same either way, metered tokens somewhere; the only question is whether the migration happens on your schedule or the platform's. One warning that applies everywhere: shared key lists on GitHub or Discord and unofficial "free DeepSeek proxy" endpoints are stolen credentials or logging middleboxes often enough that the sensible policy is categorical. Do not send anything through them.

The self-host route: MIT weights, real hardware.

DeepSeek open-sourced both V4 variants under MIT with the April 24, 2026 preview release, weights public on Hugging Face. Legally unrestricted, genuinely free of per-token fees, and full data locality: for infra teams with the hardware, this is the strongest answer on the page. The hardware is the filter. V4 Flash, the smaller variant, is 284B total parameters with 13B active; served at usable speed that is multi-GPU territory, and V4 Pro at 1.6T total is cluster territory. Below that scale, quantized community models on a single consumer GPU are the realistic local option, at a real quality distance from V4. The arithmetic rarely favors buying hardware to avoid rates this low: years of a hobby project's metered usage costs less than one serving GPU.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Picking the right route.

Whichever route starts the project, the numbers that end the search are the same: published V4 rates low enough that "free vs paid" is usually a difference of single-digit dollars per month, and a starting credit that makes the first thousands of requests cost nothing while you find that out.

• Evaluating V4 for a real project: the starting credit, because it runs on the exact ids you would ship with and the usage log prices your prompts for you.
• Sampling out of curiosity: a rotating free variant is fine; expect the cap mid-session.
• Roleplay frontends needing a custom OpenAI-compatible endpoint: the starting credit; the free-variant caps are exactly the ones that end long chats abruptly.
• Infra teams with multi-GPU hardware and privacy requirements: the MIT weights.
• Production traffic of any kind: metered access, chosen deliberately. No free mechanism is designed to carry it.
• Comparing V4 against GLM or Kimi before choosing a family: the same credit runs on glm-5, glm-5.2, and kimi-k2.6 through the same endpoint, so one evaluation budget covers the whole value tier.

FAQ