What makes it different from other AI models?

It stands out due to its open-source nature, cost-effective training methods, and use of a Mixture of Experts (MoE) model. It also incorporates chain-of-thought reasoning to enhance problem-solving.

How does it train its AI models efficiently?

It optimizes computational resources through: Optimized data processing: Reducing redundant calculations. Reinforcement learning: Enhancing decision-making abilities over time. Parallel computing: Accelerating training while maintaining accuracy.

What is the Mixture of Experts (MoE) approach?

MoE allows this ai model to divide its system into specialized sub-models (experts) that handle different tasks. It dynamically selects the appropriate expert for each input, improving efficiency while reducing computational costs.

How does it implement chain-of-thought reasoning?

It processes information step by step instead of generating responses in a single pass. This technique makes it highly effective in handling complex tasks like: Mathematical computations Programming tasks Logical deductions

AI AI Foundations

Deepseek: Features, Pricing & Accessibility

Sena Sezer

updated on Oct 30, 2025

See our ethical norms

A Chinese hedge fund spent $294,000 training an AI model that beats OpenAI’s O1 on reasoning benchmarks. Then they open-sourced it.

DeepSeek isn’t your typical AI startup. High-Flyer, an $8 billion quantitative hedge fund, funds the entire operation. No venture capital. No fundraising rounds. Just a hedge fund that decided to build AI models and give them away for free.

The results challenge assumptions about the costs of AI development. While competitors reportedly spend tens of millions training frontier models with 16,000+ GPUs, DeepSeek trained R1 using 2,048 GPUs. The kicker: R1 outperforms OpenAI’s O1 on AIME 2024 (96.3% vs 79.2%) and matches it on MATH-500¹.

Key Facts

Founded by: Liang Wenfeng (also founded High-Flyer hedge fund)
Funding: Self-funded by High-Flyer, no external investors
License: MIT (you can use it commercially, modify it, whatever)
Latest models: V3.2-Exp (March 2025), R1 (January 2025)
Training cost: $294,000 for R1 (OpenAI hasn’t disclosed O1’s training cost, but estimates run into millions)

Latest Model: DeepSeek V3 & R1

DeepSeek V3-0324 Highlights

Performance Improvements DeepSeek V3-0324 represents a significant leap forward, achieving top rankings on critical benchmarks:

MMLU-Pro: Advanced reasoning capabilities
GPQA Diamond: Scientific question answering
AIME 2024: Mathematical problem solving
LiveCodeBench: Real-world coding performance

The model demonstrates competitive performance with Claude 3.5 Sonnet across various evaluation metrics.

Technical Specifications

Model Size: ~641GB (full precision)
License: MIT (fully open-source)
Distribution: Available via Hugging Face
Quantization Options:
- 2.71-bit: Optimal balance of performance and efficiency
- 1.78-bit: Maximum compression (with quality trade-offs)

Background and Funding

DeepSeek was founded by Liang Wenfeng, whose previous venture was High-Flyer, a quantitative hedge fund valued at $8 billion and ranked among the top four in China. Unlike many AI startups that rely on external investments, DeepSeek is fully funded by High-Flyer and has no immediate plans for fundraising.

Models and Pricing

(1) The deepseek-chat model has been upgraded to DeepSeek-V3. deepseek-reasoner points to the new model DeepSeek-R1.
(2) CoT (Chain of Thought) is the reasoning content deepseek-reasoner gives the final answer before output.
(3) If max_tokens is not specified, the default maximum output length is 4K. Please adjust max_tokensto support longer outputs.
(4) Please check DeepSeek Context Caching for the details of Context Caching.
(5) The form shows the original price and the discounted price. From now until 2025-02-08 16:00 (UTC), all users can enjoy the discounted prices of API. After that, it will recover to full price.
(6) The output token count of deepseek-reasoner includes all tokens from CoT and the final answer, and they are priced equally².

Five Technical Features Worth Understanding

1. Dual-Mode Inference: Think vs Non-Think

Most models give you one answer. R1 lets you choose:

Think mode: The model shows its reasoning process. You see it break down the problem, consider alternatives, catch its own mistakes. Slower, but more accurate for complex tasks.

Non-Think mode: Direct answer, no reasoning shown. Faster, cheaper, fine for simple queries.

Toggle this in the chat interface or set it via API. Think mode costs more (you’re paying for those reasoning tokens) but catches errors that direct inference misses.

Example use case: For a “what’s 2+2” query, use Non-Think. For “design a database schema for a social network with 100M users,” use Think.

2. Sparse Attention for Long Context (V3.2-Exp)

Traditional transformers process every token against every other token. For a 100K token conversation, that’s 10 billion token-pair comparisons. Expensive.

DeepSeek Sparse Attention (DSA) processes attention selectively:

Identifies which tokens actually matter for context
Skips irrelevant token pairs
Reduces memory and compute requirements

Real impact: You can have longer conversations without costs exploding. A 100K token context that might cost $5.50 on GPT-4 costs $0.90 on DeepSeek V3.2-Exp.

3. Open Source Under MIT License

You can download R1’s weights from Hugging Face and:

Run it locally (if you have 640GB of RAM)
Modify the architecture
Use it commercially
Train on top of it
Not pay DeepSeek anything

This differs from “open weights” models with restrictive licenses. MIT means actual open source.

Catch: The newest V3.2-Exp is only available via API for now. R1 and earlier versions are downloadable.

4. Mixture-of-Experts Architecture

DeepSeek-MoE divides the model into specialized sub-networks (experts). For each query:

A router decides which experts to activate
Only 2-4 experts process the input (not all 256)
Reduces compute while maintaining quality

Think of it like a hospital: instead of every doctor examining every patient, triage routes patients to specialists.

The tradeoff: Training MoE models is harder. Inference is cheaper. DeepSeek chose to invest upfront in training to reduce ongoing costs.

5. Multi-head Latent Attention (MLA)

Technical explanation: MLA compresses the key-value cache using latent representations, reducing memory footprint by ~75% versus standard attention.

Practical explanation: Normally, as conversations get longer, memory requirements explode. MLA keeps memory usage reasonable even in 128K token conversations.

This matters if you’re running models locally or trying to serve many users simultaneously. Less memory = cheaper hardware requirements.

How to Access DeepSeek

Web interface: chat.deepseek.com (free, no login required for basic use)

Mobile apps:

iOS: Search “DeepSeek” in App Store
Android: Available on Google Play

API:

Signup at platform.deepseek.com
Get API key
Standard OpenAI-compatible endpoints
Python SDK available

Self-hosting:

Download weights from Hugging Face
Requires: 640GB RAM for full precision, or 128GB for quantized versions
VLLM and other inference frameworks supported

DeepSeek vs GPT-5: Actual Differences

Real Limitations You Should Know

The Model Is Huge

641GB at full precision. Even quantized to 2.71 bits, you need 128GB of RAM. Most developers use the API instead of self-hosting.

If you want to run it locally:

Rent a high-memory cloud instance ($5-10/hour)
Build a workstation with 256GB+ RAM ($5,000+)
Use quantized versions and accept quality degradation

Think Mode Is Slow

Chain-of-thought reasoning takes time. A query that GPT-4 answers in 2 seconds might take R1’s Think mode 10-15 seconds.

For chatbots or real-time applications, this matters. For batch processing or complex analysis, the accuracy gains justify the wait.

Content Moderation Is Opaque

Research from 2025 showed DeepSeek sometimes reasons about sensitive topics internally but removes that reasoning from the final output. You see the conclusion but not the thinking that led there.

Example: Ask about certain political topics, and R1’s internal reasoning (visible in Think mode during testing) shows one analysis but outputs a different answer.

This isn’t unique to DeepSeek—all Chinese AI companies must comply with local regulations—but it affects transparency.

Documentation Assumes Technical Knowledge

The API docs are clear if you’ve used OpenAI’s API before. If you haven’t, expect a learning curve. Community resources are growing but still limited compared to GPT ecosystem.

No Multimodal Capabilities Yet

R1 and V3.2-Exp handle text only. No images, no audio, no video. DeepSeek has a separate model (Janus) for vision tasks, but it’s not integrated into the main reasoning models.

Results of DeepSeek-R1-Lite-Preview Across Benchmarks

DeepSeek-R1-Lite-Preview achieved strong results across benchmarks, particularly in mathematical reasoning. Its performance improves with extended reasoning steps.

Source: DeepSeek³