Latest10 Sep 2026Open-weights LLM
DeepSeek-V4.1-Flash
The newest DeepSeek model and the first of its V4.1 architecture family, with image understanding built in. It replaced V4-Flash as the default API model at lower prices, and DeepSeek says its long-context memory cache needs about a quarter of the high-bandwidth memory of the previous generation. Weights and a technical report are on Hugging Face.
- 552B total MoE; 8B active for input, 16B for output (lab-stated)
- 1M-token context; up to 384K output
- Text and image input, text output
- KV cache about 1/4 the HBM of the previous generation (lab claim)
- Reasoning effort adjustable
AccessAPI (model name deepseek-flash; old V4-Flash names route to it); open weights on Hugging Face, MIT licence
Price$0.15 input (cache miss) / $0.60 output per 1M tokens off-peak; $0.30 / $1.20 peak
21 Aug 2026Open-weights LLM
DeepSeek-V4-Flash-Vision-Exp
An experimental vision version of V4-Flash, DeepSeek's first multimodal model in the V4 family. DeepSeek says it matches V4-Flash on text tasks while improving a lot on agent tasks that need to read screens and images. It launched with a free Files API for re-using uploaded images.
- Image plus text input, text output
- Text ability on par with V4-Flash (lab claim)
- Images billed at up to 384 tokens each at V4-Flash rates
- Supports Chat Completions, Messages and Responses APIs
AccessAPI (deepseek-v4-flash-vision-exp, now routed to V4.1-Flash); open weights on Hugging Face, MIT licence
13 Aug 2026Frontier LLM
DeepSeek-V4-Pro (general availability, V4-Pro-0813)
The official, non-preview release of DeepSeek's largest model, with stronger agent behaviour, three reasoning-effort levels and native support for OpenAI's Responses API so it can drive Codex. It also brought off-peak pricing at half the peak rate. DeepSeek briefly planned to retire V4-Pro in September but kept it on after user demand.
- 1.6T total / 49B active parameters (V4 announcement)
- 1M-token context; up to 384K output
- Reasoning effort: low, high, max
- SWE-Bench Verified 80.6 for V4-Pro-Max (lab model card)
AccessWeb/app via Expert Mode; API (deepseek-v4-pro); updated open weights DeepSeek-V4-Pro-0813 on Hugging Face, MIT licence
Price$0.66 input (cache miss) / $1.98 output per 1M tokens off-peak; $1.32 / $3.96 peak
24 Apr 2026Frontier LLM
DeepSeek-V4 (Pro and Flash, preview)
DeepSeek's fourth-generation family, released as a preview in two sizes. V4-Pro was billed as the strongest open model for agentic coding and maths, and V4-Flash as a cheaper model close to it in reasoning. Both made 1M-token context the default across DeepSeek's services.
- V4-Pro: 1.6T total / 49B active
- V4-Flash: 284B total / 13B active
- 1M-token context default
- Thinking and non-thinking modes
- New attention: token-wise compression plus DeepSeek Sparse Attention
Accesschat.deepseek.com (Expert / Instant mode); API with OpenAI and Anthropic formats; open weights on Hugging Face, MIT licence
27 Jan 2026Embedding/other
DeepSeek-OCR 2
A small open model that reads documents and images and turns them into text or Markdown. It introduces a 'Visual Causal Flow' encoder meant to read pages in a more human-like order, following the original DeepSeek-OCR from October 2025. It is released under the permissive Apache-2.0 licence and runs on vLLM or Transformers.
- About 3B parameters
- Document-to-Markdown and plain OCR modes
- Dynamic resolution input
- Runs on vLLM and Transformers
AccessOpen weights on Hugging Face and GitHub, Apache-2.0
1 Dec 2025Open-weights LLM
DeepSeek-V3.2 and V3.2-Speciale
The final release of the V3.2 generation, adding DeepSeek Sparse Attention for cheaper long contexts and better tool use. A separate high-compute version, V3.2-Speciale, was built only for deep reasoning; DeepSeek says it reached gold-medal level at the 2025 maths and informatics olympiads.
- 685B parameters
- DeepSeek Sparse Attention (DSA)
- Speciale: gold-level IMO 2025 and IOI 2025 (lab claim)
- Speciale has no tool calling
AccessApp, web and API (V3.2 behind deepseek-chat / deepseek-reasoner); V3.2-Speciale on a temporary API endpoint to 15 Dec 2025; open weights on Hugging Face, MIT licence
27 Nov 2025Reasoning LLM
DeepSeekMath-V2
An open maths model that checks its own proofs with a built-in verifier rather than only chasing final answers. DeepSeek reports gold-level scores on IMO 2025 and CMO 2024 and 118 of 120 on Putnam 2024, putting olympiad-level proof writing in an open model.
- 685B parameters, built on DeepSeek-V3.2-Exp-Base
- Self-verifying proof generation
- Putnam 2024: 118/120 (lab-reported)
- Gold-level IMO 2025 and CMO 2024 (lab-reported)
AccessOpen weights on Hugging Face, Apache-2.0
20 Oct 2025Embedding/other
DeepSeek-OCR
A small open vision model that studies 'optical compression': storing long text as images so a language model can read more with fewer tokens. It doubles as a fast document OCR tool and was added to the vLLM inference engine three days after release.
- About 3B parameters
- Several resolution modes (Tiny to Large, plus dynamic)
- About 2,500 tokens/s on one A100-40G (lab claim)
- PDF and image to text or Markdown
AccessOpen weights on Hugging Face and GitHub, MIT licence