DeepSeek API

DeepSeek V4 Flash Vision Experimental is an experimental multimodal version of DeepSeek's V4 Flash model, released by DeepSeek in August 2026. It is a sparse Mixture-of-Experts model with 284B total parameters and 13B activated per token, and it accepts images (JPEG, PNG, GIF, WebP) alongside text through the same API used for V4 Flash, at the same per-token pricing. DeepSeek reports that it matches V4 Flash on text tasks including agentic reasoning and world knowledge, while adding image understanding for document and chart parsing, visual question answering, and agent workflows that interleave text and images. On multimodal agent benchmarks, DeepSeek's own numbers show gains over the text-only V4 Flash, with Terminal Bench 2.1 rising from 82.7 to 83.9, ApexBench from 26.2 to 36.5, and Agents' Last Exam from 25.2 to 27.3, approaching Anthropic's Opus 4.8 (85.0 on Terminal Bench 2.1). On NL2Repo it trails Opus 4.8, 57.7 versus 69.7.

DeepSeek V4 Pro 0813 is the production release of DeepSeek's V4 Pro flagship model, a 1.6-trillion-parameter Mixture-of-Experts model with 49B parameters active per token, marking the end of a four-month preview. It keeps the preview's 1,048,576-token context window, 384,000-token max output, and three reasoning modes (non-thinking, high, and max effort). DeepSeek reports large gains over the preview build: Terminal Bench 2.1 rose from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7. At max reasoning effort it scores 80.6% on SWE-bench Verified, matching Gemini 3.1 Pro, along with 90.1% on GPQA Diamond, 87.5% on MMLU-Pro, and a Codeforces rating of 3,206, ahead of GPT-5.4's 3,168. Priced at $0.38 per million input tokens and $1.14 per million output tokens. Pinning the 0813 route locks an application to this exact checkpoint rather than the rolling deepseek-v4-pro alias, useful for agentic coding and long-context reasoning tasks.

DeepSeek V4 Flash 0731 is the July 31, 2026 dated snapshot of DeepSeek's V4 Flash, a sparse Mixture-of-Experts chat model with a 1,048,576-token context window, hosted here on Together AI's infrastructure. It keeps the April preview's architecture but was re-post-trained with a pipeline focused on coding, reasoning, agents, and tool use. In DeepSeek's own benchmarks the agentic gains over the preview are large: Terminal Bench 2.1 rises from 61.8 to 82.7, Cybergym from 38.7 to 76.7, and DeepSWE from 7.3 to 54.4. On these vendor-reported numbers it also outscores the V4 Pro preview, though third-party reproduction was not available at release. Pinning the dated 0731 route locks an application to this exact checkpoint, while a rolling deepseek-v4-flash route can move to newer versions. It fits high-volume coding assistants, chat systems, and agent workflows where cost and latency matter, and this route's 1M-token max output gives more headroom for long agent traces than the original preview route.

DeepSeek V4 Flash is a lightweight, efficiency-focused Mixture-of-Experts model from DeepSeek, with 284B total parameters and 13B activated per token. It supports a 1M-token context window and configurable reasoning modes (standard, high, and max thinking effort). Designed as the fast and economical option in the V4 family, Flash delivers reasoning capabilities that closely approach the larger V4 Pro, and performs on par with it on simpler agentic tasks. In its max reasoning mode, it achieves comparable reasoning scores to Pro when given a larger thinking budget. At $0.14/M input and $0.28/M output tokens, it's one of the cheapest frontier-tier models available — well suited for high-throughput workloads like coding assistants, chat systems, and agent pipelines where latency and cost matter most.

DeepSeek V4 Pro is a 1.6T-parameter Mixture-of-Experts model from DeepSeek with 49B parameters activated per token, supporting a 1M-token context window. It is positioned as the strongest open-weight model currently available. V4 Pro leads all open-source models in math, coding, and STEM reasoning. On LiveCodeBench it scores 93.5, ahead of Gemini 3.1 Pro (91.7) and Claude Opus 4.6 (88.8). Its Codeforces rating of 3206 also tops GPT-5.4 (3168). On agentic tool-use benchmarks like MCPAtlas, it reaches near-parity with Opus 4.6. DeepSeek acknowledges it trails GPT-5.4 and Gemini 3.1 Pro overall by roughly 3–6 months of frontier development. Priced at $1.74/M input and $3.48/M output — a fraction of comparable closed-source models — it's a strong pick for complex reasoning, agentic coding, and knowledge-intensive tasks.

DeepSeek V3.2 is the December 2025 flagship model featuring DeepSeek Sparse Attention for efficiency and massive reinforcement learning post-training, achieving GPT-5-level performance. It's the first DeepSeek model to integrate thinking directly into tool-use and excels at agentic AI tasks.

DeepSeek V3.2-Speciale is a high-compute variant designed exclusively for maximum reasoning accuracy, achieving gold-medal performance in IMO 2025, IOI 2025, and ICPC World Finals. It rivals Gemini 3.0 Pro but requires higher token usage and doesn't support tool calling.

DeepSeek OCR is a vision model from DeepSeek, released in October 2025, built for optical character recognition and document parsing. It takes images as input through a chat-style API and returns text, including documents converted to markdown. Its core idea is optical context compression, representing text as a small number of vision tokens and decoding them back to text. In DeepSeek's paper it reaches 97% decoding precision at compression ratios under 10x, and about 60% at 20x. On OmniDocBench it outperforms GOT-OCR 2.0 while using only 100 vision tokens per page (versus 256), and beats MinerU 2.0 with fewer than 800 vision tokens (versus 6,000+). It fits document-heavy workloads: parsing PDFs and scans, extracting text and tables, and generating training data for other models.

DeepSeek V3.2-Exp is the September 2025 experimental predecessor to V3.2, introducing DeepSeek Sparse Attention architecture through continued training on V3.1-Terminus. It served as a testing ground for the sparse attention innovations later refined in V3.2.

DeepSeek V3.1-Terminus is the September 2025 refined update to V3.1, addressing user-reported issues like language mixing and improving Code Agent and Search Agent capabilities. It represents the final, most stable version of the V3 architecture before V3.2.

DeepSeek V3.1 is an August 2025 hybrid model that combines the capabilities of V3 and R1, supporting both thinking and non-thinking modes via chat template switching. It features 671B parameters (37B activated), 128K context, and significantly improved tool-calling and agent capabilities.

DeepSeek R1-0528 is the May 2025 major update to R1, featuring dramatically improved reasoning depth with nearly double the thinking tokens (23K vs 12K average) and approaching performance of O3 and Gemini 2.5 Pro. It adds function calling support, reduced hallucinations, and improved AIME accuracy from 70% to 87.5%.

DeepSeek V3-0324 is the March 2025 update to DeepSeek V3, incorporating reinforcement learning techniques from R1 to significantly improve reasoning, coding, and frontend development capabilities. It became the first open-source model to outperform all proprietary non-reasoning models on benchmarks, exceeding GPT-4.5 in math and coding tasks.

DeepSeek R1 Distill Qwen 32B is a 32 billion parameter dense model fine-tuned from Qwen 2.5 using R1-generated reasoning data, achieving state-of-the-art results for dense models. It outperforms OpenAI o1-mini on various benchmarks while being efficient enough for local deployment.

DeepSeek R1 Distill Llama 70B is a 70 billion parameter dense model fine-tuned from Llama 3.3-70B-Instruct using 800K reasoning samples generated by DeepSeek R1. It brings R1's reasoning capabilities to a more accessible size while maintaining strong performance on math and coding benchmarks.

DeepSeek Reasoner is the API alias for DeepSeek's reasoning models (R1 series), which use chain-of-thought reasoning to solve complex math, coding, and logic problems. It displays its thinking process before arriving at answers and achieves performance comparable to OpenAI o1.

DeepSeek R1 is DeepSeek's first-generation reasoning model released January 2025, trained via large-scale reinforcement learning to achieve performance comparable to OpenAI o1 on math, code, and reasoning tasks. It pioneered open-source reasoning capabilities with self-verification and reflection behaviors.

DeepSeek Chat is the general-purpose conversational alias that points to the latest DeepSeek V3 chat model, a 671B parameter Mixture-of-Experts LLM optimized for everyday conversations, coding assistance, and general tasks. It supports 128K context and provides fast, direct responses without explicit reasoning chains.