DeepSeek
Hangzhou-based AI lab publishing open-weights language models and technical reports
Market Metrics
V4-Pro (Apr 2026)
1.6T / 49B active
open-weights preview; 1M context
V3 Training Run (disclosed)
2.788M H800 GPU-hours / $5.576M
at an assumed $2 per GPU-hour; excludes prior research and ablation experiments
Product Specifications
| Product | Type | Status |
|---|---|---|
| DeepSeek-V4 | Open-weights MoE (V4-Pro 1.6T/49B, V4-Flash 284B/13B, 1M context) | Preview released (Apr 2026) |
| DeepSeek-V3 | 671B parameter MoE open-weights base model (37B active) | Released (Dec 2024) |
| DeepSeek-R1 | Open-weights reasoning model (chain-of-thought) | Released (Jan 2025) |
| DeepSeek-Coder family | Open-weights code models (1.3B to 33B) | Released |
| DeepSeek Chat | Free chatbot on web, App Store and Google Play | GA |
| DeepSeek API | Developer API compatible with OpenAI ChatCompletions and Anthropic formats | GA |
Partnerships & Investors
DeepSeek-AI's Fire-Flyer 2 paper credits High-Flyer Quant and describes a cluster of 10,000 PCIe A100 GPUs.
Baidu AI Cloud's Qianfan platform listed DeepSeek-R1 and V3 on 5 Feb 2025; ByteDance's Volcano Ark lists deepseek-v3-241226 and deepseek-r1-250120; Huawei Cloud's MaaS lists DeepSeek-V3.2, V3.1 and R1-0528.
Key Milestones
DeepSeek LLM
Open-source 7B and 67B parameter Base and Chat models trained on 2 trillion tokens
DeepSeek-V2
236B parameter MoE (21B active) with a 128K context window; introduces Multi-head Latent Attention (MLA) and DeepSeekMoE
DeepSeek-V3
671B parameter MoE (37B active) trained on 14.8T tokens; technical report puts the official training run at 2.788M H800 GPU-hours, $5.576M at an assumed $2 per GPU-hour, excluding prior research and ablation experiments
DeepSeek-R1
Open-weights reasoning model released under the MIT License; DeepSeek says its performance is on par with OpenAI-o1, and publishes six distilled models
Italy Limits DeepSeek Data Processing
Italy's data protection authority orders, as a matter of urgency, a limitation on processing of Italian users' data by Hangzhou DeepSeek Artificial Intelligence and Beijing DeepSeek Artificial Intelligence, and opens an investigation
DeepSeek-V4 Preview
V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), both open-sourced with a 1M-token context window; DeepSeek says V4-Pro rivals the world's top closed-source models, and the V4 report validates its expert-parallel scheme on NVIDIA GPUs and Huawei Ascend NPUs
KPI data tables
DeepSeek operating metrics behind the KPI charts, newest period first (up to 24 periods per chart). Periods use each company's own fiscal labels where the source does; each table names its source.
| Period | DeepSeek V4-Pro Max | OpenAI GPT-5.4 xHigh | Google Gemini-3.1-Pro High |
|---|---|---|---|
| Codeforces rating | 3,206 | 3,168 | 3,052 |
Source: DeepSeek-V4-Pro model card, DeepSeek-V4-Pro-Max vs Frontier Models table (huggingface.co/deepseek-ai/DeepSeek-V4-Pro); DeepSeek-V4 Technical Report (arxiv.org/abs/2606.19348); DeepSeek API Docs, DeepSeek-V4 Preview Release (24 Apr 2026).
| Period | DeepSeek V4-Pro Max | Claude Opus-4.6 Max | Google Gemini-3.1-Pro High |
|---|---|---|---|
| LiveCodeBench | 93.5 | 88.8 | 91.7 |
| SWE-bench Verified (%) | 80.6 | 80.8 | 80.6 |
Source: DeepSeek-V4-Pro model card, DeepSeek-V4-Pro-Max vs Frontier Models table (huggingface.co/deepseek-ai/DeepSeek-V4-Pro); DeepSeek-V4 Technical Report (arxiv.org/abs/2606.19348); DeepSeek API Docs, DeepSeek-V4 Preview Release (24 Apr 2026).
Frequently asked questions
What does DeepSeek do?
Hangzhou-based lab publishing open-weights models and technical reports (DeepSeek V2, V3, R1 and the V4 preview of Apr 2026). DeepSeek's V3 technical report puts the official training run at 2.788M H800 GPU-hours, $5.576M at an assumed $2 per GPU-hour, excluding prior research and ablation experiments; its R1 paper puts R1 at $294K (147K H800 GPU-hours). DeepSeek-R1 (Jan 2025) was widely cited as the inflection point in the open-versus-closed competitive map. DeepSeek is privately held and reports no financials; it discloses no funding round or valuation. Its API is priced per token (V4-Pro $1.32 input and $3.96 output per million tokens at peak rates, half that off-peak). DeepSeek was founded in 2023 and is headquartered in Hangzhou, China.
Who are DeepSeek's main competitors?
Sterling groups DeepSeek in the AI Software industry, alongside OpenAI (private), Anthropic (private), SpaceXAI (private), Mistral AI (private), Cohere (private) and AI21 Labs (private).
If I'm investing in DeepSeek, what similar companies should I consider?
If you are researching DeepSeek, also look at Alibaba (Qwen), Tencent (Hunyuan) and Baidu (Ernie). You can compare all of them side by side or browse the full AI Software industry on Sterling.