兔子API 兔子API
中

model

DeepSeek

Direct answerThe current official IDs are deepseek-v4-flash and deepseek-v4-pro; both enable thinking by default and can disable it. This gateway supports only names and paths explicitly enabled in its model list and protocol guide.

Updated · Reviewed

Current family

As of this review, the current DeepSeek API IDs are deepseek-v4-flash and deepseek-v4-pro. The official page lists a 1M context, 384K maximum output, JSON mode, and tool calling. Limits, prices, and concurrency can change, so production configuration should read current official documentation instead of copying a static table indefinitely.

Thinking is not a separate fixed family

Both V4 models enable thinking by default. Direct calls can switch thinking.type=enabled|disabled and tune reasoning_effort. Do not keep assuming that chat always means non-reasoning and reasoner always means reasoning. A mode suffix shown by this gateway is a convenience alias or channel mapping, not a new official ID.

Direct versus gateway model names

Direct access uses https://api.deepseek.com, a DeepSeek API key, and an official model ID. This gateway uses its own base URL, Bearer token, and an enabled gateway model name. An official ID is not callable here until a channel enables or maps it. This model page declares only /v1/chat/completions; official Responses, Anthropic-compatible, and beta routes are not gateway endpoints by implication.

Streaming, JSON, and tools

Official Chat streams use SSE and terminate with [DONE]; supported stream_options.include_usage requests aggregate usage. JSON mode needs an explicit JSON instruction or it may stream whitespace. Tool arguments must be validated and executed by the application, with allowlists, authorization, and human confirmation for high-impact actions.

Selection and cost safety

Compare Flash and Pro on the same production samples for accuracy, latency, output tokens, and tool success. Try disabled thinking for simple work and enable it for multi-step math, code, or dense constraints. Set output caps, timeouts, concurrency, balance alerts, and bounded retries; handle 401, 402, 429, and 500/503 by category rather than blindly replaying requests.

Reproducibility

Retain the full request, returned model, usage, request ID, and date, together with regression samples. Provider implementations, prices, and concurrency can change, so audited workloads should revalidate tools, JSON, context, and output boundaries before upgrades.

Use cases

  • Deep reasoning
  • Coding
  • Long context

API protocols

  • /v1/chat/completions

FAQ

What are the current official DeepSeek model IDs?

As of this review, the official IDs are deepseek-v4-flash and deepseek-v4-pro. Both list a 1M context, 384K maximum output, and switchable thinking. These limits and prices can change, so recheck the official pricing page before launch.

Do standard generation and reasoning still require two different models?

Current V4 models enable thinking by default. Direct calls can set thinking.type=disabled and tune reasoning_effort. If this gateway lists convenience names with mode suffixes, they are gateway mappings rather than additional official model IDs.

Why can the official model work while the gateway says it does not exist?

The provider catalog and gateway model list are separate. Direct access uses an official ID and provider key; gateway calls require a name enabled by the administrator and a gateway token, with any provider mapping handled by the channel.

How do I control thinking latency and cost?

Evaluate disabled thinking first for simple extraction or classification and enable it for harder reasoning. Always cap output, timeouts, concurrency, and balance exposure. Streams end with [DONE], and supported stream_options.include_usage can record actual consumption.

  • DeepSeekDeepSeek's current official model IDs are deepseek-v4-flash and deepseek-v4-pro; thinking is enabled by default and can be disabled. Direct calls use a DeepSeek API key, while this gateway uses its own Bearer token.
  • OpenAI-compatible API guideThis site's OpenAI-compatible routes use a Bearer token; Responses, Chat Completions, and media endpoints have distinct objects and streaming semantics and must match the target model.
  • Choosing and calling coding modelsSelect coding models by patch correctness, test pass rate, tool success, latency, and cost on real repositories, not marketplace copy or a single benchmark.
  • AI API quickstart: from API key to first requestAn AI API call needs a matching base URL, API key, model ID, and endpoint; start with minimal curl, then add error handling, cost controls, and key security.
  • Common AI API error codes and troubleshootingHandle AI API failures by parsing the HTTP status, provider code, error type, message, retry hints, and request ID together. Fix request, authentication, billing, and hard-quota causes first; retry only transient throttling, network failures, and selected 5xx errors with strict bounds.

Official sources

  1. DeepSeek Models and Pricing Official
  2. DeepSeek Chat Completions API Official
  3. DeepSeek Error Codes Official