DeepSeek models
Chinese AI lab that develops the open-weights DeepSeek model family, including the V-series general models and R-series reasoning models, and runs the DeepSeek API and app.
Hangzhou, Chinadeepseek.com
- Releases
- 38
- Open weights
- 37
- Latest
- Oct 17, 2026
October 2026
1 releaseDeepSeek-V4.1-Flash-1017
DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic, and visual results while shortening outputs and task runtime. It replaced the previous version on the DeepSeek web, app, and API.
- Modalities
- text, image → text
- Context
- 1M context
- Parameters
- MoE: 552B backbone, 8B active (prefill), 16B active (decode)
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Served through the deepseek-flash API model name with unchanged pricing and calling method
- Trained with reinforcement learning with self-feedback for end-to-end project development
- Weights on Hugging Face and ModelScope under the MIT License
September 2026
1 releaseDeepSeek-V4.1-Flash
DeepSeek-V4.1-Flash is the smallest model in DeepSeek's new architecture family, a multimodal mixture-of-experts model with native image understanding. It uses a Causal Encoder-Decoder design that activates fewer parameters for input than for output.
- Modalities
- text, image → text
- Context
- 1M context
- Parameters
- MoE: 552B backbone, 8B active (prefill), 16B active (decode)
- License
- MIT
- At launch
- API, Open weights
- Pre-trained from scratch on a 45T-token multimodal corpus
- Global KV cache about 4 times smaller than DeepSeek-V4-Flash
- Replaced V4-Flash and V4-Flash-Vision-Exp; API model name deepseek-flash
- Supports a continuously controllable reasoning effort from 1 to 100
August 2026
2 releasesDeepSeek-V4-Flash-Vision-Exp
DeepSeek-V4-Flash-Vision-Exp is DeepSeek's first experimental multimodal model in the V4 family, adding visual modules to the V4-Flash architecture. DeepSeek says it matches V4-Flash on text tasks while improving multimodal agent performance.
- Modalities
- text, image → text
- License
- MIT
- At launch
- API, Open weights
- API model name deepseek-v4-flash-vision-exp, supporting mixed text and image input
- Launched alongside a new Files API for reusing uploaded images
- Retired on the API on September 10, 2026, when V4.1-Flash replaced it
- Weights published on Hugging Face under the MIT License
DeepSeek-V4-Pro (GA, 0813)
DeepSeek-V4-Pro-0813 is the general availability release of DeepSeek-V4-Pro, superseding the April preview with greatly enhanced agentic capabilities. It keeps the preview's model structure and adds an attached DSpark speculative decoding module.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.6T total, 49B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Rolled out on the DeepSeek app, web, and API under the unchanged deepseek-v4-pro model name
- Added low, high, and max reasoning effort levels for V4-Pro and V4-Flash
- DeepSeek API gained native OpenAI Responses API support with this release
July 2026
1 releaseDeepSeek-V4-Flash (0731)
DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the April preview with substantially enhanced agentic capabilities. It keeps the preview's architecture and size and was re-post-trained, with a speculative decoding module attached.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 284B total, 13B active
- License
- MIT
- At launch
- API, Open weights
- Released to the deepseek-v4-flash API as a public beta; the app, web, and V4-Pro API were unchanged
- Supports low, high, and max reasoning effort levels
- Weights on Hugging Face as DeepSeek-V4-Flash-0731 under the MIT License
April 2026
1 releaseDeepSeek-V4 (Preview)
DeepSeek-V4 is a preview release of two open-weights mixture-of-experts models, V4-Pro and V4-Flash. It introduced a hybrid attention design combining token-wise compression with DeepSeek Sparse Attention and made 1M-token context the default across DeepSeek services.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.6T total, 49B active (V4-Pro); 284B total, 13B active (V4-Flash)
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Both models pre-trained on more than 32T tokens and support thinking and non-thinking modes
- API supports OpenAI ChatCompletions and Anthropic formats via deepseek-v4-pro and deepseek-v4-flash
- Legacy deepseek-chat and deepseek-reasoner names were scheduled for retirement on July 24, 2026
January 2026
1 releaseDeepSeek-OCR 2
DeepSeek-OCR 2 is DeepSeek's second document OCR vision-language model, introducing a visual encoding approach the lab calls Visual Causal Flow. It follows DeepSeek-OCR and is released with vLLM and Transformers inference code.
- Modalities
- text, image → text
- License
- Apache-2.0
- At launch
- Open weights
- Subtitle: Visual Causal Flow, aimed at more human-like visual encoding
- Weights on Hugging Face under the Apache 2.0 license
- Announced in the DeepSeek-OCR repository on January 27, 2026
December 2025
2 releasesDeepSeek-V3.2
DeepSeek-V3.2 is the official successor to V3.2-Exp, a reasoning-first model built for agents. It is DeepSeek's first model to integrate thinking directly into tool use, supporting tool calls in both thinking and non-thinking modes.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Built on DeepSeek Sparse Attention for efficient long-context inference
- Trained with agent data synthesized across more than 1,800 environments and 85,000 complex instructions
- Powers both deepseek-chat (non-thinking) and deepseek-reasoner (thinking) on the API
DeepSeek-V3.2-Speciale
DeepSeek-V3.2-Speciale is a high-compute reasoning variant of DeepSeek-V3.2 that pushes maximum reasoning capability at the cost of higher token usage. DeepSeek reports gold-level results at IMO, CMO, the ICPC World Finals, and IOI 2025.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Open weights
- Served through a temporary API endpoint available until December 15, 2025
- No tool calling support at launch, offered for community evaluation and research
- Weights on Hugging Face under the MIT License
November 2025
1 releaseDeepSeekMath-V2
DeepSeekMath-V2 is an open-weights model for self-verifiable mathematical reasoning and theorem proving, built on DeepSeek-V3.2-Exp-Base. DeepSeek reports gold-level scores on IMO 2025 and CMO 2024 with scaled test-time compute.
- Modalities
- text → text
- License
- Apache-2.0
- At launch
- Open weights
- Targets rigorous step-by-step proofs rather than final-answer accuracy alone
- Built on top of DeepSeek-V3.2-Exp-Base
- Weights released under the Apache 2.0 license
October 2025
1 releaseDeepSeek-OCR
DeepSeek-OCR is an open-weights vision-language model that explores contexts optical compression, studying the role of vision encoders from an LLM-centric viewpoint. It converts document images to text and markdown.
- Modalities
- text, image → text
- License
- MIT
- At launch
- Open weights
- Released on GitHub and Hugging Face on October 20, 2025
- Supported in upstream vLLM from October 23, 2025
- Weights under the MIT License
September 2025
2 releasesDeepSeek-V3.2-Exp
DeepSeek-V3.2-Exp is an experimental model built on V3.1-Terminus that debuts DeepSeek Sparse Attention for more efficient long-context training and inference. DeepSeek says it performs on par with V3.1-Terminus.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Live on the DeepSeek app, web, and API at launch
- V3.1-Terminus stayed available on a temporary API for comparison until October 15, 2025
- Released with GPU kernels in TileLang and CUDA and a technical report
DeepSeek-V3.1-Terminus
DeepSeek-V3.1-Terminus is an update to DeepSeek-V3.1 that addresses user feedback on language consistency and strengthens its code agent and search agent performance. It replaced V3.1 across the app, web, and API.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Fewer Chinese and English mix-ups and fewer abnormal characters
- Upgraded both deepseek-chat (non-thinking) and deepseek-reasoner (thinking)
- Weights on Hugging Face
August 2025
1 releaseDeepSeek-V3.1
DeepSeek-V3.1 is a hybrid model that supports thinking and non-thinking modes in one model, with stronger tool use and multi-step agent skills from post-training. DeepSeek called it its first step toward the agent era.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 671B total, 37B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- V3.1 Base adds 840B tokens of continued pre-training for long-context extension on top of V3
- 128K context for both deepseek-chat and deepseek-reasoner
- API added Anthropic format support and strict function calling in beta
May 2025
1 releaseDeepSeek-R1-0528
DeepSeek-R1-0528 is an update to DeepSeek-R1 with improved reasoning, enhanced front-end generation, and reduced hallucinations. It added JSON output and function calling support to the reasoning model.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Replaced R1 behind the deepseek-reasoner API model with no change to API usage
- Released with DeepSeek-R1-0528-Qwen3-8B, a distilled 8B model
- Weights on Hugging Face under the MIT License
April 2025
1 releaseDeepSeek-Prover-V2
DeepSeek-Prover-V2 is an open-weights model for formal theorem proving in Lean 4, trained with cold-start data from a recursive proof-search pipeline powered by DeepSeek-V3. It comes in 7B and 671B sizes.
- Modalities
- text → text
- Parameters
- 7B, 671B
- License
- DeepSeek Model License
- At launch
- Open weights
- DeepSeek-Prover-V2-671B is trained on top of DeepSeek-V3-Base
- DeepSeek-Prover-V2-7B builds on Prover-V1.5-Base with context extended to 32K tokens
- Use is subject to the DeepSeek model license
March 2025
1 releaseDeepSeek-V3-0324
DeepSeek-V3-0324 is an updated DeepSeek-V3 checkpoint with a major boost in reasoning, stronger front-end development skills, and smarter tool use. It also moved the V3 weights to the MIT License.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Replaced V3 behind the deepseek-chat API model on March 24, 2025
- DeepSeek recommends it for non-complex reasoning tasks with DeepThink turned off
- Weights on Hugging Face under the MIT License
January 2025
3 releasesJanus-Pro
Janus-Pro is an advanced version of Janus, DeepSeek's autoregressive framework that unifies multimodal understanding and image generation by decoupling visual encoding into separate pathways. It significantly improves both understanding and visual generation.
- Modalities
- text, image → text, image
- Parameters
- 1B, 7B
- License
- DeepSeek Model License
- At launch
- Open weights
- Built on DeepSeek-LLM-1.5b-base and DeepSeek-LLM-7b-base
- Single unified transformer for image understanding and text-to-image generation
- Code under the MIT License; model use under the DeepSeek Model License
DeepSeek-R1
DeepSeek-R1 is an open-weights reasoning model trained with large-scale reinforcement learning in post-training, released with a technical report. DeepSeek also released DeepSeek-R1-Zero, trained with reinforcement learning without supervised fine-tuning.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 671B total, 37B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Code and weights under the MIT License; API outputs may be used for fine-tuning and distillation
- Available in the API as deepseek-reasoner and as DeepThink on chat.deepseek.com
- 128K context length per the model card
DeepSeek-R1-Distill
DeepSeek-R1-Distill is a set of six dense models fine-tuned on reasoning samples generated by DeepSeek-R1, based on Qwen2.5 and Llama 3 checkpoints. They bring R1-style reasoning to sizes from 1.5B to 70B parameters.
- Modalities
- text → text
- Parameters
- 1.5B, 7B, 8B, 14B, 32B, 70B
- License
- MIT
- At launch
- Open weights
- Fine-tuned with 800K samples curated with DeepSeek-R1
- Qwen-based models derive from Qwen2.5 (Apache 2.0); Llama-based models from Llama 3.1 8B and Llama 3.3 70B
- Released the same day as DeepSeek-R1
December 2024
3 releasesDeepSeek-V3
DeepSeek-V3 is an open-weights mixture-of-experts model built on Multi-head Latent Attention and the DeepSeekMoE architecture, pre-trained on 14.8T tokens. It replaced the model behind the deepseek-chat API and ran about 3 times faster than V2.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 671B total, 37B active
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- 671B total parameters with 37B activated per token
- Generates about 60 tokens per second, per DeepSeek
- Code under the MIT License; weights under the DeepSeek Model License, which permits commercial use
DeepSeek-VL2
DeepSeek-VL2 is a series of mixture-of-experts vision-language models that improves on DeepSeek-VL in visual question answering, OCR, document, table, and chart understanding, and visual grounding. It comes in Tiny, Small, and standard variants.
- Modalities
- text, image → text
- Parameters
- MoE: 1.0B, 2.8B, 4.5B active (Tiny, Small, VL2)
- License
- DeepSeek Model License
- At launch
- Open weights
- Three variants with 1.0B, 2.8B, and 4.5B activated parameters
- Uses a dynamic tiling strategy for high-resolution images
- Model use is subject to the DeepSeek Model License, which supports commercial use
DeepSeek-V2.5-1210
DeepSeek-V2.5-1210 is the final update in the DeepSeek-V2.5 series, improving math, coding, writing, and roleplay. DeepSeek released it alongside Internet Search on its web chat and described it as the close of the V2 line.
- Modalities
- text → text
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- Upgraded the deepseek-chat API model
- Optimized file upload and webpage summarization experiences
- Weights on Hugging Face as DeepSeek-V2.5-1210
November 2024
2 releasesDeepSeek-R1-Lite-Preview
DeepSeek-R1-Lite-Preview is a reasoning model preview that shows its thought process in real time. It launched on chat.deepseek.com, with DeepSeek saying open-source models and an API would follow.
- Modalities
- text → text
- At launch
- Consumer app
- Available on DeepSeek's web chat at launch, without API access
- Showed steadily improving results as thought length increased, per DeepSeek
- Preceded the open-weights DeepSeek-R1 release in January 2025
JanusFlow
JanusFlow is a unified model for image understanding and generation that integrates an autoregressive language model with rectified flow. DeepSeek reports that rectified flow can be trained within the language model framework without complex architectural changes.
- Modalities
- text, image → text, image
- Parameters
- 1.3B
- License
- DeepSeek Model License
- At launch
- Open weights
- Released as JanusFlow-1.3B on Hugging Face
- Part of DeepSeek's Janus series of unified multimodal models
- Model use under the DeepSeek Model License
October 2024
1 releaseJanus
Janus is an autoregressive framework that unifies multimodal understanding and image generation by decoupling visual encoding into separate pathways while using a single transformer. DeepSeek released it as Janus-1.3B.
- Modalities
- text, image → text, image
- Parameters
- 1.3B
- License
- DeepSeek Model License
- At launch
- Open weights
- Separate visual encoding paths for understanding and generation
- Single unified transformer architecture
- Code under the MIT License; model use under the DeepSeek Model License
September 2024
1 releaseDeepSeek-V2.5
DeepSeek-V2.5 merges DeepSeek-V2-Chat-0628 and DeepSeek-Coder-V2-0724 into one model that combines general conversation and coding abilities. It is better aligned with human preferences and improves writing and instruction following.
- Modalities
- text → text
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- API users can reach it through either deepseek-chat or deepseek-coder
- Strengthened resistance to jailbreak attacks while reducing over-refusal
- Weights on Hugging Face
August 2024
1 releaseDeepSeek-Prover-V1.5
DeepSeek-Prover-V1.5 is a 7B open-weights model for theorem proving in Lean 4, pre-trained on DeepSeekMath-Base and refined with reinforcement learning from proof assistant feedback. It improves on DeepSeek-Prover-V1, whose weights were also published.
- Modalities
- text → text
- Parameters
- 7B
- License
- DeepSeek Model License
- At launch
- Open weights
- Released as Base, SFT, and RL models
- Uses reinforcement learning from proof assistant feedback (RLPAF)
- Model use subject to the DeepSeek model license, which supports commercial use
July 2024
1 releaseDeepSeek-Coder-V2-0724
DeepSeek-Coder-V2-0724 is an updated DeepSeek-Coder-V2 checkpoint with improved general capabilities through alignment optimization. It upgraded the deepseek-coder API model and was later merged into DeepSeek-V2.5.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 236B total, 21B active
- License
- DeepSeek Model License
- At launch
- API, Open weights
- Upgraded the deepseek-coder API model on July 24, 2024
- Weights published on Hugging Face as DeepSeek-Coder-V2-Instruct-0724
- One of the two models merged to create DeepSeek-V2.5
June 2024
2 releasesDeepSeek-V2-Chat-0628
DeepSeek-V2-Chat-0628 is an upgraded DeepSeek-V2-Chat whose base model was replaced with the Coder-V2 base, improving code generation, math, and reasoning. It also strengthened role-playing.
- Modalities
- text → text
- License
- DeepSeek Model License
- At launch
- API, Open weights
- Upgraded the deepseek-chat API model on June 28, 2024
- Weights published on Hugging Face as DeepSeek-V2-Chat-0628
- Later merged with DeepSeek-Coder-V2-0724 to form DeepSeek-V2.5
DeepSeek-Coder-V2
DeepSeek-Coder-V2 is an open-weights mixture-of-experts code model further pre-trained from an intermediate DeepSeek-V2 checkpoint with an additional 6T tokens. It substantially improved coding and math reasoning while keeping general language performance.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 236B total, 21B active; Lite 16B total, 2.4B active
- License
- DeepSeek Model License
- At launch
- API, Open weights
- Released in 236B and 16B Lite sizes, each with Base and Instruct models
- 128K context length
- Upgraded the deepseek-coder API model on June 14, 2024
May 2024
2 releasesDeepSeek-V2-Lite
DeepSeek-V2-Lite is a smaller mixture-of-experts model in the DeepSeek-V2 family, released as a base model and an SFT chat model. It uses the same economical training and efficient inference architecture as DeepSeek-V2.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- MoE: 16B total, 2.4B active
- License
- DeepSeek Model License
- At launch
- Open weights
- 16B total parameters with 2.4B activated
- 32K context length
- Released as DeepSeek-V2-Lite and DeepSeek-V2-Lite-Chat
DeepSeek-V2
DeepSeek-V2 is a mixture-of-experts language model designed for economical training and efficient inference, pre-trained on 8.1T tokens. Compared with DeepSeek 67B, it saves 42.5% of training costs and reduces the KV cache by 93.3%.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 236B total, 21B active
- License
- DeepSeek Model License
- At launch
- API, Consumer app, Open weights
- 236B total parameters with 21B activated per token
- 128K context length
- Released as a base model and an RL-tuned chat model
March 2024
1 releaseDeepSeek-VL
DeepSeek-VL is a vision-language model family released in 1.3B and 7B sizes with base and chat variants. The 7B model uses a hybrid SigLIP and SAM vision encoder that supports 1024 by 1024 image input.
- Modalities
- text, image → text
- Parameters
- 1.3B, 7B
- License
- DeepSeek Model License
- At launch
- Open weights
- Four models: 1.3B and 7B, each as base and chat
- DeepSeek-VL-7B-base is trained on around 400B vision-language tokens
- Supports commercial use under the DeepSeek Model License
February 2024
1 releaseDeepSeekMath 7B
DeepSeekMath 7B is a math-specialized model initialized from DeepSeek-Coder-v1.5 7B and further pre-trained on 500B tokens of math-related, natural language, and code data. Its RL variant was trained with DeepSeek's Group Relative Policy Optimization (GRPO).
- Modalities
- text → text
- Parameters
- 7B
- License
- DeepSeek Model License
- At launch
- Open weights
- Released as Base, Instruct, and RL models
- Introduced the GRPO reinforcement learning algorithm
- Supports commercial use under the DeepSeek Model License
January 2024
1 releaseDeepSeekMoE 16B
DeepSeekMoE 16B is a mixture-of-experts language model that uses fine-grained expert segmentation and shared expert isolation. Trained on 2T English and Chinese tokens, it matches DeepSeek 7B with about 40% of the computation.
- Modalities
- text → text
- Parameters
- MoE: 16.4B total
- License
- DeepSeek Model License
- At launch
- Open weights
- 16.4B total parameters
- Deployable on a single 40GB GPU without quantization
- Released as Base and Chat models
November 2023
2 releasesDeepSeek LLM
DeepSeek LLM is a general-purpose language model family released in 7B and 67B sizes as base and chat models. The models were trained from scratch on 2T English and Chinese tokens.
- Modalities
- text → text
- Context
- 4K context
- Parameters
- 7B, 67B
- License
- DeepSeek Model License
- At launch
- Open weights
- Pre-trained with a 4,096-token sequence length
- DeepSeek also released intermediate base model checkpoints for research
- Supports commercial use under the DeepSeek Model License
DeepSeek Coder
DeepSeek Coder is a series of open-weights code models trained from scratch on 2T tokens of 87% code and 13% natural language. Base models support project-level code completion and infilling.
- Modalities
- text → text
- Context
- 16K context
- Parameters
- 1.3B, 5.7B, 6.7B, 33B
- License
- DeepSeek Model License
- At launch
- Open weights
- Sizes of 1.3B, 5.7B, 6.7B, and 33B, with Base and Instruct models
- Pre-trained with a 16K window and a fill-in-the-blank task
- Covers more than 80 programming languages; supports commercial use
Cite and reuse this data
These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.
Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".
Citation
The AI Engineering Editors. "DeepSeek models: release history and timeline." The AI Engineering, updated Oct 17, 2026. https://theaiengineering.com/releases/deepseek
HTML link
<a href="https://theaiengineering.com/releases/deepseek">DeepSeek models: release history and timeline</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering
Other labs
Google106
Alibaba Qwen84
OpenAI79Mistral AI59
Meta44
- Z.ai43
Microsoft42Tencent42
- ByteDance Seed40
- Liquid AI34
Cohere31- IBM30
xAI30- NVIDIA29
MiniMax28
Stability AI28- Anthropic26
Ai223StepFun22
Amazon21- Baidu21
- ElevenLabs21
Black Forest Labs18
- Kuaishou16
- Meituan16
Runway16- Moonshot AI15
Midjourney14Salesforce13
- Xiaomi13
Nous Research11- Perplexity10
AI21 Labs8- Apple8
Ideogram8
01.AI7
Hugging Face7
Lightricks7Luma AI7
Reka7- Suno7
- Databricks5
Leonardo.Ai5- Pika5
Recraft5
Inflection AI3MBZUAI IFM3
- Snowflake3
Thinking Machines Lab3
TypeSafe AI2- Cognition1
Genmo1Reflection AI1
Ricursive Intelligence1