DeepSeek models

DeepSeek's open MoE models: R1 made open reasoning models mainstream, and DeepSeek-V4 brings 1M-token context under the MIT licence.

Open weights
Models 6
Context up to 160K
Updated Nov 2026

Hugging Face Compare

On this page

At a glance
Models & versions
VRAM & download size
Run it locally
Best uses
Strengths & limits
About DeepSeek
Guides
FAQ
Compare

At a glance

Type: Open-weight models
Developer: DeepSeek (China)
First release: Nov 2023
Models tracked: 6 (2 current)
Longest context: 160K tokens
Abilities: Tool callingReasoning modeImage input
Licence: MIT licence for R1, V3.1 and V4, so commercial use is allowed. The original DeepSeek-V3 weights used DeepSeek's own model licence.
Smallest local option: DeepSeek-R1 1.5B · ~2 GB
Website: deepseek.com
Hugging Face: huggingface.co/deepseek-ai

Newest model

DeepSeek-V4.1 Flash

Released Sep 2026

Smallest option

DeepSeek-R1 1.5B

~2 GB VRAM · 1.1 GB download

DeepSeek models and versions

Model	Released	Sizes	Context	Abilities	Licence	Status
DeepSeek-V4.1 Flash Multimodal MoE with 1M context and a compressed KV cache	Sep 2026	552B Too large for most servers · Ollama cloud	—	ToolsReasoningVision	MIT	Current
DeepSeek-V4 Pro Flagship with three reasoning levels and 1M context	Apr 2026	1.6T 1.6T total, 49B active (MoE) Too large for most servers · Ollama cloud	—	ToolsReasoning	MIT	Current
DeepSeek-OCR Token-efficient OCR for documents	Oct 2025	3B	8K	Vision	MIT	Previous generation
DeepSeek-V3.1 Hybrid model with thinking and non-thinking modes	Aug 2025	671B 671B total, 37B active (MoE)	160K	ToolsReasoning	MIT	Previous generation
DeepSeek-R1 Open reasoning model; small sizes are Qwen and Llama distills	Jan 2025	1.5B · 7B · 8B · 14B · 32B · 70B · 671B 671B (+ 1.5B–70B distills)	160K	ToolsReasoning	MIT	Previous generation
DeepSeek-V3 The MoE model that put DeepSeek on the map	Dec 2024	671B 671B total, 37B active (MoE)	160K	Text only	DeepSeek Model License	Legacy

Sizes, context windows and abilities come from the Ollama library and Hugging Face model cards, checked 14 Nov 2026. Release months are the first public upload. Spotted a mistake? Tell us.

How much VRAM does DeepSeek need?

Download sizes per Ollama build and the GPU memory each needs with an 8K context. On a CPU-only server the same number is the RAM you need. Longer contexts need more memory; work it out with the LLM VRAM calculator.

Model	4-bit (default)	8-bit	16-bit	VRAM (4-bit)	Fits on
DeepSeek-OCR 3B	6.7 GB	—	6.7 GB	~8 GB	8 GB GPU
DeepSeek-V3.1 671B	404 GB	713 GB	1300 GB	~445 GB	Multi-GPU
DeepSeek-R1 1.5B	1.1 GB	1.9 GB	3.6 GB	~2 GB	8 GB GPU
DeepSeek-R1 7B	4.7 GB	8.1 GB	15 GB	~6 GB	8 GB GPU
DeepSeek-R1 8B	5.2 GB	8.9 GB	16 GB	~6.5 GB	8 GB GPU
DeepSeek-R1 14B	9 GB	16 GB	30 GB	~10.5 GB	12 GB GPU
DeepSeek-R1 32B	20 GB	35 GB	66 GB	~22.5 GB	24 GB GPU
DeepSeek-R1 70B	43 GB	75 GB	141 GB	~48 GB	48 GB GPU
DeepSeek-R1 671B	404 GB	713 GB	1300 GB	~445 GB	Multi-GPU

VRAM = 4-bit download + about 10% runtime overhead + 0.5 GB for an 8K context, rounded up. Mixture-of-experts (MoE) models still need all their weights in memory, but they run much faster than a dense model of the same size, so they also work well split between GPU and system RAM.

Run DeepSeek-OCR locally

# Install Ollama (Linux), then pull and chat
curl -fsSL https://ollama.com/install.sh | sh
ollama run deepseek-ocr:3b

New to Ollama? Follow Install Ollama on Ubuntu 26.04 (or 24.04) for GPU drivers, the systemd service and remote access.

# Ollama serves an OpenAI-compatible API on port 11434
curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-ocr:3b", "messages": [{"role": "user", "content": "Hello"}]}'

# Full-precision weights from Hugging Face, OpenAI-compatible API on :8000
vllm serve deepseek-ai/DeepSeek-OCR

vLLM loads the original 16-bit weights, so it needs roughly the 16-bit size above in GPU memory, plus room for the context. See Install vLLM on Ubuntu 26.04.

What DeepSeek is good for

Reasoning on modest hardware

DeepSeek-R1 distills (7B, 8B, 14B) show their chain of thought and fit on 8–12 GB GPUs.

Self-hosted frontier model

V4 Pro and V4.1 Flash are MIT-licensed, so companies can host them on their own GPU clusters.

Document OCR

DeepSeek-OCR turns scans into text with few tokens per page.

Strengths and limits

Strengths

MIT licence on current models: no user caps or usage rules.
1M-token context on V4 and V4.1.
Very efficient MoE designs; the hosted API is among the cheapest.

Watch out for

Full V3/V4 models need hundreds of GB of memory; only the R1 distills run on one consumer GPU.
The small R1 sizes are fine-tuned Qwen and Llama models, not the real R1.
Some organisations restrict Chinese-developed models; check your policy.

About DeepSeek

DeepSeek is a Hangzhou-based lab funded by the quant fund High-Flyer. DeepSeek-V3 (December 2024), a 671B mixture-of-experts model with 37B active parameters, showed that a frontier-class model could be trained cheaply. DeepSeek-R1 (January 2025) followed with open reasoning under the MIT licence, along with small “distilled” versions built on Qwen and Llama that anyone could run.

DeepSeek-V3.1 (August 2025) merged thinking and non-thinking modes into one model. DeepSeek-V4 Pro (April 2026) scaled to 1.6T parameters (49B active) with a 1M-token context and three thinking levels. DeepSeek-V4.1 Flash (September 2026) adds image input and a new encoder-decoder design that shrinks the KV cache, the memory that grows with context length.

On a single server, the realistic options are the R1 distills; the full models need a GPU cluster or a hosted endpoint such as Ollama cloud.

Guides for running DeepSeek

Tool LLM VRAM Calculator: How Much GPU Memory Do You Need? Updated Nov 2026 AI & LLMs How to Install Ollama on Ubuntu 26.04 and Run LLMs Locally Updated Nov 2026 AI & LLMs How to Install llama.cpp on Ubuntu 26.04 (CPU and NVIDIA GPU) Updated Nov 2026 AI & LLMs How to Install vLLM on Ubuntu 26.04 (OpenAI-Compatible LLM Server) Updated Nov 2026 AI & LLMs How to Install Open WebUI on Ubuntu 26.04 (Self-Hosted ChatGPT Alternative) Updated Nov 2026

Frequently asked questions

What is the latest DeepSeek model?

The newest DeepSeek model is DeepSeek-V4.1 Flash, released in Sep 2026.

How much VRAM do I need to run DeepSeek locally?

The smallest option, DeepSeek-R1 1.5B, is a 1.1 GB download at 4-bit and needs about 2 GB of GPU memory (or RAM on a CPU-only machine) with an 8K context. Larger sizes need more; see the hardware table on this page, and use the VRAM calculator for longer contexts.

How do I run DeepSeek with Ollama?

Install Ollama, then run "ollama run deepseek-r1:1.5b". Ollama downloads the 4-bit build and starts a chat; the same model is also available on the local API at http://localhost:11434.

Can I use DeepSeek commercially?

MIT licence for R1, V3.1 and V4, so commercial use is allowed. The original DeepSeek-V3 weights used DeepSeek's own model licence. Always read the licence of the exact model you deploy; it is linked in the models table.

Is "ollama run deepseek-r1" the real DeepSeek-R1?

Only the 671B tag is. The 1.5B–70B tags are smaller Qwen and Llama models fine-tuned on R1's reasoning (distills). They are useful, but much weaker than the full model.

Compare DeepSeek

← All AI models