DeepSeek Models: Full List, Features & Use Cases

DeepSeek's model lineup has grown from a small open-source coding model into one of the most closely watched open-weight AI families in the world, known for near-frontier benchmark scores at a fraction of the price of closed competitors. This guide lists every major DeepSeek model in order from the newest to the oldest, with a concise overview of each model's specs, key features, and intended use — plus links to try the ones available on Chat Smith.

Released April 24, 2026 under the MIT license, DeepSeek V4 Pro (1.6T total / 49B active parameters) and DeepSeek V4 Flash (284B / 13B active) both ship with a 1M-token context window, made economical by a new hybrid attention design. Pricing: V4 Pro at $0.435/1M input, $0.87/1M output; V4 Flash at $0.14/1M input, $0.28/1M output. Best for: agentic coding and long-context reasoning (Pro), high-volume latency-sensitive tasks (Flash).

Released December 1, 2025, DeepSeek V3.2 was the first DeepSeek model to integrate thinking directly into tool use, scoring 73.1% on SWE-Bench Verified. Its high-compute sibling, V3.2-Speciale, pushed reasoning further, posting gold-medal-level olympiad math results and 96% on AIME. The DeepSeek Reasoner V3.2 endpoint exposes this reasoning behavior through a dedicated "thinking" mode. Best for: general Q&A and agent tasks (V3.2), competition-level math and long-form reasoning (Speciale).

Released September 29, 2025, V3.2-Exp introduced DeepSeek Sparse Attention (DSA), cutting API pricing roughly in half versus the prior generation while improving throughput on long-context tasks. It served as the experimental preview that shaped the full V3.2 release two months later. Best for: reference only — superseded by V3.2.

Released August 21, 2025, DeepSeek V3.1 merged "thinking" and regular response modes into a single model for the first time, removing the need to pick separate chat and reasoning endpoints. The September 22 Terminus refinement cleaned up bilingual (English/Chinese) output, strengthened agentic tool use, and made long-context handling cheaper. Best for: reference only — superseded by V3.2.

Released May 28, 2025, R1-0528 was a major update to DeepSeek's original reasoning model, jumping to 91 on AIME, 73 on LiveCodeBench, and 57.6% on SWE-bench Verified. DeepSeek also shipped an 8B distilled version based on Qwen3 small enough to run on a laptop. Best for: reference only — reasoning capability now lives inside the V3.2/V4 lineup.

Released March 24, 2025, V3-0324 was a quiet but significant update to the original V3 base model — a 685B-parameter MoE model released on Hugging Face under the MIT license, and the foundation later releases built on. Best for: reference only — fully superseded by V3.1 and later.

Released January 20, 2025, DeepSeek R1 is the release that made the company a household name, wiping hundreds of billions of dollars off US tech stocks the week it launched. It demonstrated that pure reinforcement-learning-driven reasoning, without heavy supervised fine-tuning, could produce a model competitive with OpenAI's o1. The experimental R1-Zero checkpoint tested RL from scratch, while R1 added cold-start data to improve readability. Six smaller distilled models let developers run a lighter version locally. Best for: reference only — succeeded by R1-0528 and the V3.2/V4 lineup.

Released December 25–26, 2024, DeepSeek V3 scaled DeepSeek's Mixture-of-Experts design to 671 billion total parameters (37B active), trained on 14.8 trillion tokens for a reported cost far below what Western labs were spending on comparable models. It's the base architecture that R1 was later built on. Best for: reference and research — self-hostable under a permissive license.

Released September 5, 2024 and revised that December, V2.5 merged DeepSeek's general-purpose chat model with its coding-focused line into a single model, simplifying the lineup ahead of V3. Best for: historical reference only.

Released May 6, 2024, DeepSeek V2 (236B total / 21B active parameters, 128K context) is the architectural starting point for every later DeepSeek flagship: it introduced Multi-head Latent Attention (MLA) for cheaper KV caching alongside a Mixture-of-Experts design, cutting training cost by 42.5% and KV cache size by 93.3% versus DeepSeek's earlier dense models. DeepSeek Coder V2 followed in June, extending the architecture to 338 programming languages. Best for: historical reference — the architecture, not the checkpoint, is what matters today.

Released in November 2023, DeepSeek's first models — DeepSeek Coder and the 7B/67B DeepSeek LLM — were trained on roughly 2 trillion tokens and were competitive with Meta's LLaMA-2 70B on code, math, and reasoning benchmarks at launch. They put DeepSeek on the map as a serious open-source lab well before V3 or R1 made global headlines. Best for: historical reference only.

DeepSeek's lineup now spans open-weight general models, dedicated reasoners, and a fast/pro split at the flagship tier — all released under permissive licenses that undercut closed competitors by an order of magnitude on price. For teams and individuals who want to use these models without self-hosting, Chat Smith gives you access to the available DeepSeek models alongside GPT, Claude, Gemini, and Grok — all from a single interface. Browse the full AI model catalog on Chat Smith to find the right model for your workflow.