DeepSeek V4 Flash API

Yes. DeepSeek V4 Flash is live on the EvoLink production route. Send requests with the model ID deepseek-v4-flash on Chat Completions or Responses, and check the model ID, pricing, context, and supported workflows on this page before scaling up.

Yes. The model ID you send in the request is deepseek-v4-flash — the same string as this page URL. The current GA version is the 0731 build (promoted to stable on July 31, 2026), and the unchanged ID serves it automatically. The old aliases deepseek-chat and deepseek-reasoner were retired upstream on July 24, 2026.

DeepSeek documents a 1,000,000-token context window. There is no separate long-context price tier — billing follows the live per-token rates and cached-input mechanics shown in the Pricing section, so test representative context sizes and cache hits instead of assuming every request should use the full window.

Text input and text output only — DeepSeek V4 Flash has no vision capability on any protocol. On the Messages route, image and document content types are rejected; on Responses, image and file attachments are converted to placeholders rather than understood. Route screenshot or document-understanding tasks to a vision-capable model on the same EvoLink gateway.

Use Chat Completions for familiar chat, streaming, and client function flows. Evaluate Responses — added with the 0731 build — for longer-running agent workflows. The current API section and EvoLink docs remain the source for exact request fields and verified route support.

Valid tiers are low, high, and max, and the default is high — medium is accepted but silently maps to high, so comparing medium against high produces no real difference. Start with low for routine work, compare low and high on the same tasks, and reserve max for the hardest problems where a failed attempt costs more than the extra reasoning tokens.

Cached input can reduce repeated-context cost, while cache misses, long outputs, retries, and repeated tool steps can make an agent loop more expensive than the headline input rate suggests. Use the live Pricing section and final request charges when calculating cost per accepted task.

There is no per-token RPM or TPM limit upstream. The constraint is an account-level concurrency cap — around 2,500 concurrent requests for the Flash tier, roughly five times the Pro cap — with 429 returned beyond it and requests dropped after about 10 minutes of queueing. That headroom is Flash’s throughput advantage: keep in-flight requests below your measured ceiling, use exponential backoff, and it handles high-concurrency batch work well.

Choose Flash for classification, summarization, short edits, and high-concurrency batch pipelines, where its speed and roughly 5x concurrency ceiling matter most. Choose Pro for agent chains longer than about eight steps and for fact-sensitive tasks where reasoning depth outweighs throughput. Both tiers share the same EvoLink API, so routing by task type is a configuration change.

Run the same real tasks with consistent context, tools, and reasoning settings. Compare result quality, response time, token mix, tool performance, and total cost to decide which EvoLink route fits each traffic class.

Yes — the Flash 0731 weights are published under the MIT license on Hugging Face. Open weights and the hosted API are independent access paths: the EvoLink route serves the hosted API, and the same MIT weights on third-party hosts are what make fallback routing possible.

Keep a model that already handles the same workload reliably and leave route selection configurable. If rate limits, timeouts, or invalid output occur, EvoLink can route to GPT, Claude, Kimi, or another suitable alternative.