FAQ

For heavy reasoning, multi-file complex coding, and hard math, choose deepseek-v4-pro (1.6T/49B, no vision, 500 concurrent requests); for high concurrency, long documents, cost-sensitive scenarios, and image input needs, choose deepseek-flash (now pointing to V4.1-Flash-1017, 552B, 2500 concurrent requests). Both offer the same 1M context / 384K output.

Thinking mode is enabled by default in the V4 series; the API controls it via the reasoning parameter. The RLSF training of V4.1-Flash-1017 makes its chain of thought more compact—official benchmarks show it outputs significantly fewer tokens than the previous generation, which is the main reason the per-task cost dropped from $0.50 to $0.20.

The API is billed by usage with no free tier: V4.1-Flash costs $0.15/$0.60 off-peak and $0.30/$1.20 at peak, with cache-hit input at $0.003; V4-Pro costs $0.66/$1.98 per million tokens off-peak. However, the weights of V4-Pro, V4-Flash, and V4.1-Flash are open-sourced under the MIT license, so you can self-deploy at zero API cost. This site's playground offers free streaming demos.

Starting 2026-09-14, the official team routes v4-pro calls in the API to V4.1-Flash and recommends that new projects use deepseek-flash directly. The old IDs have a migration window; please refer to the migration notes in the official API documentation.

Yes: the V4 series offers a 1M input window and 384K maximum output, and its hybrid attention (CSA+HCA) brings the compute overhead for 1M workloads down to about 27% of V3.2, with KV cache at roughly 10%. Third-party evaluations also note that ultra-long-range retrieval recall still trails Claude Opus 4.6 (MRCR), so pairing with retrieval is recommended for factual tasks.

Both V4-Pro and V4-Flash have their code and weights released under the MIT license on Hugging Face (Flash is about 160GB); they can be used commercially, modified, and have no regional restrictions. The V4.1-Flash series also has open weights.

Yes. The official API provides both OpenAI- and Anthropic-compatible formats, so migration usually only requires changing the base_url and model name; JSON output, tool calling, and streaming are supported.

The playground on this site (deepseek.wang) does not store conversation content and only relays it via streaming; for commercial privacy needs, we recommend the official enterprise solution or self-hosting with the MIT-licensed weights.