DeepSeek Open Source Models Complete List

Model Series Overview

DeepSeek currently open-sources the following 6 major model series, covering scenarios such as conversation, reasoning, coding, multimodal, vision, and mathematical proof.

One-Click Download of DeepSeek Open Source Models

Complete download links for all models. Ollama is suitable for quick personal experience, Hugging Face is suitable for development integration and fine-tuning.

DeepSeek R1 — Enhanced Reasoning Model

Trained with reinforcement learning, it excels in mathematics, programming, and logical reasoning. The full R1 version with 671B parameters is the strongest reasoning model; the distilled versions transfer R1's reasoning capabilities to smaller models like Qwen and Llama, allowing them to run on consumer hardware. Available in 7 sizes from 1.5B to 671B, it is the most downloaded model on Ollama.

R1 Distilled Versions: Full Specs and Ollama Downloads

The R1 distilled versions transfer R1 671B's reasoning capabilities to smaller models via knowledge distillation, allowing you to experience powerful reasoning on consumer-grade hardware. Below are all available versions:

DeepSeek Coder — Code-Specialized Models

Focused on code generation and understanding. Coder V2 uses a 236B MoE architecture, supports 338 programming languages and 128K context, and surpasses GPT-4 Turbo in coding ability. Coder V1 offers lightweight 1.3B~33B sizes and supports FIM code completion.

Coder V2 — Latest Flagship Code Model

Coder V1 — Lightweight Code Model

Trained on 2 trillion tokens (87% code + 13% Chinese and English natural language), supports FIM code completion. Suitable for single-GPU deployment.

DeepSeek VL2 — Mixture-of-Experts Vision Language Model

DeepSeek VL2 is an open-source MoE (Mixture-of-Experts) vision language model from DeepSeek, adopting a mixture-of-experts architecture to achieve strong visual understanding while maintaining efficient inference. The model supports dynamic resolution and can process images of arbitrary resolution without forced cropping or resizing.

Core Architecture Design

Dynamic Resolution Technology

Comparison of Three Sizes

VL2 offers three sizes to meet different scenario needs. All sizes are open-sourced under the MIT License:

Key Capabilities

Inference Code Example

from transformers import AutoModelForCausalLM

from deepseek_vl2.models import DeepseekVLV2Processor, DeepseekVLV2ForCausalLM
import torch

# 1. Load model and processor
model_path = "deepseek-ai/deepseek-vl2-small"
vl_chat_processor = DeepseekVLV2Processor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer

vl_gpt = DeepseekVLV2ForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
).cuda().eval()

# 2. Multi-turn conversation example (supports dynamic resolution images)
conversation = [
    {
        "role": "User",
        "content": "<image>\nPlease describe the content of this image and extract the text in it.",
        "images": ["./document.png"]
    },
    {
        "role": "Assistant",
        "content": "This is a financial report..."
    },
    {
        "role": "User",
        "content": "<image>\nWhat is the trend in this chart?",
        "images": ["./chart.png"]
    },
]

# 3. Prepare inputs and inference
prepare_inputs = vl_chat_processor(
    conversations=conversation,
    images=["./document.png", "./chart.png"],
    force_batchify=True
).to(device=vl_gpt.device, dtype=vl_gpt.dtype)

# 4. Generate answer
inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
outputs = vl_gpt.language_model.generate(
    inputs_embeds=inputs_embeds,
    attention_mask=prepare_inputs.attention_mask,
    max_new_tokens=512,
    do_sample=False,
)

answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(answer)

DeepSeek Prover V2 — Mathematical Theorem Proving

A model focused on formal mathematical theorem proving. It automates proof in the Lean 4 proof assistant, representing cutting-edge exploration of AI in mathematical reasoning.

DeepSeek Open Source Models Quick Comparison

A table to see the differences between all DeepSeek open source models at a glance, and quickly find the model you need.

DeepSeek Model Benchmark Performance Rankings

The following shows the performance of DeepSeek's core models on mainstream benchmarks. Data source: DeepSeek official technical report. Comparison of V3, R1 (full version 671B), and Coder (V2 version) scores gives you a clear view of each model's strengths.

DeepSeek Model Training Data Unveiled

The scale and composition of training data for DeepSeek models are key to understanding the boundaries of model capabilities. Below are the training data details for DeepSeek's core open-source models, all sourced from DeepSeek's official technical reports and GitHub repositories.

DeepSeek Model Open Source Community Impact

DeepSeek open-source models have had a huge impact on the global AI community. Below are key community statistics as of mid-2026, showing DeepSeek's popularity on platforms such as GitHub, Hugging Face, and Ollama.

GitHub Repository Stars

Hugging Face Platform Data

Ollama Platform Downloads

Awesome DeepSeek Integration

Community Contribution Ecosystem

Learn More

DeepSeek Model Selection Guide

Not sure which DeepSeek model to choose? The following helps you quickly find the most suitable model based on scenario, hardware, and requirements.

Select by Scenario

Select by Hardware

Choose by Needs

DeepSeek Open-Source Model FAQ