DeepSeek

DeepSeek's October 2026 checkpoint of the V4.1 Flash workhorse, refreshed for stronger agentic coding with shorter, faster answers.

DeepSeek's cost-efficient V4.1 tier and the first model built on the Causal Encoder-Decoder (CED) architecture: a 552B MoE activating 8B parameters on input and 16B on output.

Experimental vision-enabled DeepSeek V4 Flash with image understanding added on top of the 0731 snapshot.

The GA release of DeepSeek's open 1.6T-parameter MoE flagship for advanced reasoning and long-horizon agents.

The GA release of DeepSeek's efficiency-optimized V4 Flash, re-post-trained for coding, reasoning, and agent workflows.

Open-source 1.6T-parameter MoE with 49B active, built for advanced reasoning and long-horizon agents.

Efficiency-optimized DeepSeek V4 at 284B total / 13B active for fast, high-throughput inference.

Open-source model with GPT-5 class reasoning performance.

Experimental DeepSeek V3.2 variant with DeepSeek Sparse Attention.

Large hybrid reasoning model supporting thinking and non-thinking modes.

Updated DeepSeek R1 with performance on par with OpenAI o1.

March 2024 iteration of the DeepSeek V3 685B MoE model.

Open-source reasoning model matching OpenAI o1 performance.

Original DeepSeek V3 — the model that put DeepSeek on the map.

Distilled reasoning model based on Qwen 2.5 32B using DeepSeek R1 outputs.

Distilled reasoning model based on Llama 3.3 70B using DeepSeek R1 outputs.