DeepSeek: every model, side by side

DeepSeek V4.1 Flash supports text and image inputs with text output and a 1M context window. OpenRouter token pricing varies by provider and time of use.

conversation · reasoning · tool-use

Experimental vision-enabled version of DeepSeek V4 Flash 0731, adding image understanding to the base model. Supports text and image inputs with a 1M context window on OpenRouter.

conversation · reasoning · tool-use

DeepSeek's large mixture-of-experts model and the general-availability release of V4 Pro. Keeps the 1M-token context window and hybrid attention of the V4 line, aimed at reasoning, coding and long-horizon agent work.

conversation · reasoning · code-generation · analysis · agentic-tool-use

DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model from DeepSeek, with 13B active parameters out of 284B total. This re-post-trained revision of DeepSeek V4 Flash is suited for coding, reasoning, and agent workflows.

conversation · reasoning · code-generation · analysis · agentic-tool-use

DeepSeek V4 Flash, a mixture-of-experts model with 284B total and 13B activated parameters over a 1M token context. Hybrid attention handles the long context, reasoning modes are configurable, and the whole thing is tuned for throughput: coding assistants, chat and agent loops.

conversation · reasoning · code-generation · analysis · agentic-tool-use

DeepSeek V4 Pro, a mixture-of-experts model with 1.6T total and 49B activated parameters over a 1M token context. The same hybrid attention architecture as V4 Flash at much larger scale, with multiple reasoning modes and strong knowledge, mathematics and software engineering scores.

conversation · reasoning · code-generation · analysis · agentic-tool-use

The high-compute variant of DeepSeek V3.2, with reinforcement learning post-training scaled past the base model. It keeps DeepSeek Sparse Attention for long context, and DeepSeek reports it ahead of GPT-5 on hard reasoning workloads and comparable to Gemini 3.0 Pro, with coding and tool use intact.

conversation · reasoning · code-generation · analysis · tool-use

DeepSeek V3.2 introduces DeepSeek Sparse Attention, cutting training and inference cost without giving up long-context quality. Reinforcement learning post-training puts its reported reasoning in the GPT-5 class, with gold-medal results on the 2025 IMO and IOI, plus better tool-use compliance.

conversation · reasoning · code-generation · analysis · tool-use

DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention (DSA) for efficient long-context. Reasoning toggle supported via boolean flag.

conversation · reasoning · code-generation · analysis

DeepSeek V3.1 runs one set of weights in two modes: deepseek-chat answers directly, deepseek-reasoner thinks first. Open weights on Hugging Face, 160K context, tuned for agentic tool calling.

conversation · reasoning · code-generation · analysis · agentic-tool-use · function-calling · tool-use

DeepSeek R1 0528 is the May 28th update to the original DeepSeek R1. Performance on par with OpenAI o1, but open-sourced and with fully open reasoning tokens. It's 671B parameters in size, with 37B active in an inference pass. Fully open-source model.

conversation · reasoning · code-generation · analysis

A 671B parameter model, speculated to be geared towards logic and mathematics. Likely an upgrade from DeepSeek-Prover-V1.5. Released on Hugging Face without an announcement or description.

reasoning · analysis · conversation · code-generation

DeepSeek R1 is a reasoning model developed entirely via reinforcement learning, offering cost efficiency at $0.14/million tokens vs. OpenAI o1's $15, with strong code generation and analysis capabilities.

conversation · reasoning · code-generation · analysis

DeepSeek V3 (March 2024) shows significant improvements in reasoning capabilities with enhanced MMLU-Pro (81.2%), GPQA (68.4%), AIME (59.4%), and LiveCodeBench (49.2%) scores. Features improved front-end web development, Chinese writing proficiency, and function calling accuracy.

conversation · reasoning · web-design · code-generation · analysis