The complete DeepSeek catalog: current flagship models available via API and previous-generation classics, all compatible with the OpenAI / Anthropic SDK.
Data updated October 2026 · The official recommendation is to use deepseek-flash directly for new projects
The latest fast tier available via the API (invoked as deepseek-flash): RLSF self-feedback reinforcement learning, DeepSWE v1.1 77.0%, approx. $0.20 per task; thinking mode enabled by default, 2500 concurrent requests.
Flagship with 1.6T total parameters / 49B active: SWE-Bench Verified 80.6%, LiveCodeBench 93.5%, Terminal-Bench 2.0 67.9%, Codeforces 3206; 500 concurrent requests, no vision support.
552B MoE, released 09-10: native image input, approaching V4-Pro (0813) across multiple benchmarks; output price cut by about 70% vs V4-Flash, with cache hits at $0.003.
284B / 13B activated: Putnam-200 Pass@8 81.0%, SWE-Bench 79.0%; weights around 160GB, open under MIT.
The mainstay before V4: pioneer of the sparse attention approach; V4's 1M efficiency is benchmarked against it (27% FLOPs / 10% KV).
The open-source reasoning model that shook the industry in 2025 and the starting point of the pure RL training route; its reasoning paradigm has since been folded into the V4 series' thinking mode.
A representative of the code-specialized era, retired along with V4's general coding capabilities (LiveCodeBench 93.5%).
deepseek-flash (V4.1-Flash-1017): Vision + 1M context + RLSF agentic capabilities, off-peak $0.15/$0.60, 2500 concurrency—the best value-for-money option for most applications.
deepseek-v4-pro: The 1.6T flagship for multi-file complex coding and hard reasoning; no vision, 500 concurrency, suited for offline batch processing and highly demanding tasks.
The weights of V4-Pro / V4-Flash / V4.1-Flash are all openly available on Hugging Face under the MIT license, and can be used commercially and deployed on internal networks.
大模型网摘