Name: DeepSeek v3
SKU: 23082

DeepSeek v3 is a significant open-source LLM based on 671 B MoE parameters, boasting comparable performance to GPT-4 and Claude at much lower expense.

Introduction

DeepSeek v3 came out in late 2024 as an entirely open‑source model under MIT license. It uses a Mixture‑of‑Experts (MoE) structure that engages 37 B parameters per token, producing GPT‑4‑level performance on benchmarks such as AlpacaEval 2.0 and Model Arena‑Hard. It was trained on 14.8 T tokens with ~2,000 Nvidia H800 GPUs for ~55 days for ~$5.6 M. It’s about 3× speedier than V2, offering 60 tokens/sec inference.

You can use it through DeepSeek’s free chat UI, self-host through Hugging Face, or embed via API. API costs are approximately $0.27/1 M input tokens and $1.10/1 M output tokens, 50% discount during off-peak hours.

Open‑Source

Cost‑Effective

High‑Performance

Multi‑Modal

Review

DeepSeek v3 is a significant open-source LLM based on 671 B MoE parameters, boasting comparable performance to GPT-4 and Claude at much lower expense. It includes support for a 64 K token context window, function calling and JSON output. It’s quicker than v2 and adaptable enough for chat application, coding and reasoning. It’s affordable yet powerful as a substitute for proprietary models.

Features

Open‑source LLM

MIT-licensed code and weights, complete transparency and flexibility.

MoE architecture

671 B parameters with 37 B active, MLA and multi‑token prediction for scalable performance.

Large context support

Supports up to 64 K tokens, ideal for long documents.

High inference speed

60 tokens/sec, roughly 3× as fast as v2.

API with tool‑use

Function calling, JSON output, prefix and FIM completion.

Flexible access

Through free web chat, locally through Hugging Face, or API for integration.

Best Suited for

Developers

code generation, debugging, multimodal reasoning with cost-effective backbone.

Researchers & analysts

process long documents with large context support.

employ high-end AI without the high cost.

AI fans

check out MoE, LLMs, open-source models.

Strengths

9× lower cost compared to GPT‑4 and Claude for token consumption.

Leading of open-source tops, ~85.5 win rate against Alpaca, equal to closed models.

MIT license for extensive usage and research.

60 tokens/sec makes it feasible for real-time applications.

Weakness

MoE requires additional infrastructure to host locally.

US flagged connections to Chinese military, censorship of sensitive themes.

Getting started with: step by step guide

Getting started with DeepSeek V3 is easy:

Step 1: Use the Web Chat

Proceed to Deep‑Seek.chat and initiate conversation without signup with model dropdown to select V3. No installation required.

Step 2: Self‑Host via Hugging Face

Clone repository through GitHub: git clone https://github.com/deepseek-ai/DeepSeek-V3.git. Then install dependencies and download weights from Hugging Face hub.

Step 3: Integrate with API

Sign up at DeepSeek’s API portal, retrieve key, and begin using endpoints with token‑based pricing. Supports function calling, prefix/FIM and JSON outputs.

Step 4: Optimize Usage

Take advantage of off-peak pricing (50% discounts) and cache-hit billing to control costs.

Frequently Asked Questions

Q: What is DeepSeek v3?

A: A large open‑source Mixture‑of‑Experts language model rivaling GPT‑4.

Q: Is it free?

A: Yes via web chat and open‑source download. API usage is paid based on tokens.

Q: How to access it?

A: Use chat UI at Deep‑Seek.chat, self‑host through Hugging Face, or access via API.

Pricing

Token‑based model with discounts optional:

Time period	Input (cache‑miss/hit)	Output
UTC 00:30‑16:30	$0.27 / $0.07 per M	$1.10 per M
UTC 16:30‑00:30	$0.135 / $0.035 per M	$0.55 per M

Alternatives

GPT‑4o (OpenAI)

Good performance, proprietary, higher expense.

Claude 3.5 Sonnet

Great closed‑source reasoning, more expensive.