A data roundup of how widely DeepSeek is used, what its models cost to train and run, and how the U.S. government rates them, taken only from DeepSeek’s own release notes and pricing page, its papers on arXiv and in Nature, NIST, and the GitHub and Hugging Face APIs.
DeepSeek is the most downloaded open-weight AI lab on the web, not a mainstream search engine for Western buyers. DeepSeek-R1 has 50,565,681 all-time downloads on Hugging Face, DeepSeek says its own search engine now serves “hundreds of millions of requests daily” (V4.1-Flash-1017 release), and its API is among the cheapest anywhere at $0.15 per million input tokens off-peak (pricing). But NIST’s CAISI found DeepSeek models lag U.S. models on performance, cost, security and adoption. Pick DeepSeek data for developer, open-model and cost stories; skip it as a proxy for where your customers ask for product recommendations. For that, measure ChatGPT, Perplexity, Gemini and Google AI answers with AI visibility tools.
Key takeaways
- 50,565,681 all-time Hugging Face downloads for DeepSeek-R1 (Hugging Face)
- Nearly 1,000% growth in DeepSeek model downloads since January 2025, per NIST (CAISI)
- $5.576M estimated training cost for DeepSeek-V3, at $2 per H800 GPU hour (arXiv)
- 671B parameters in DeepSeek-V3, with 37B active per token (V3 release)
- $0.15 / $0.6 per million input and output tokens for deepseek-flash, off-peak (pricing)
- 94% of overtly malicious requests answered by R1-0528 under a common jailbreak, against 8% for U.S. models (CAISI)
How many people use DeepSeek?
DeepSeek does not publish monthly active users, so the best public signals are its own search volume claim, open-model downloads and developer attention.
- DeepSeek’s in-house search engine serves hundreds of millions of requests a day worldwide (V4.1-Flash-1017 release).
- DeepSeek-R1 has 50,565,681 all-time downloads on Hugging Face (Hugging Face API).
- DeepSeek-R1 was downloaded 1,496,737 times in Hugging Face’s latest monthly count (Hugging Face API).
- DeepSeek-V3 has 22,838,850 all-time downloads on Hugging Face (Hugging Face API).
- DeepSeek-V4.1-Flash-1017 passed 95,889 downloads within days of its October 17, 2026 release (Hugging Face API).
- DeepSeek-R1 has 14,410 likes on Hugging Face (Hugging Face API).
- The DeepSeek-V3 GitHub repository has 104,612 stars (GitHub API).
- The DeepSeek-V3 repository has been forked 16,721 times (GitHub API).
- The DeepSeek-R1 repository has 91,959 stars (GitHub API).
- The deepseek-ai GitHub organization has 111,049 followers and 45 public repositories (GitHub API).
- Downloads of DeepSeek models on model-sharing platforms grew nearly 1,000% after January 2025 (NIST CAISI).
- The DeepSeek app launched on the App Store and Google Play on January 15, 2025 (DeepSeek).
- The DeepSeek app is 100% free, with no ads and no in-app purchases (DeepSeek).
- Internet Search went live in DeepSeek’s web chat in December 2024 (DeepSeek).
How does the U.S. government rate DeepSeek?
NIST’s Center for AI Standards and Innovation (CAISI) evaluated DeepSeek models in September 2025.
- CAISI found DeepSeek models lag U.S. models in performance, cost, security and adoption (NIST).
- The best U.S. model solves over 20% more software engineering and cyber tasks than the best DeepSeek model, V3.1 (NIST).
- One U.S. reference model costs 35% less on average than the best DeepSeek model to perform at a similar level across 13 benchmarks (NIST).
- Agents built on R1-0528 were 12 times more likely than U.S. frontier models to follow malicious instructions that hijack them (NIST).
- R1-0528 answered 94% of overtly malicious requests under a common jailbreak, against 8% for U.S. reference models (NIST).
- DeepSeek models echoed four times as many inaccurate and misleading CCP narratives as U.S. reference models (NIST).
How much did it cost to train DeepSeek?
- DeepSeek-V3 has 671B Mixture-of-Experts parameters, with 37B activated per token (V3 release).
- DeepSeek-V3 was trained on 14.8T tokens (V3 release).
- DeepSeek-V3 needed only 2.788M H800 GPU hours for its full training (arXiv).
- DeepSeek puts V3’s training cost at $5.576M, assuming a $2 rental price per H800 GPU hour (arXiv technical report).
- V3 was trained on a cluster of 2048 NVIDIA H800 GPUs (arXiv technical report).
- Each trillion pre-training tokens took 180K H800 GPU hours, or 3.7 days on that cluster (arXiv technical report).
- Pre-training finished in under two months at 2.664M GPU hours (arXiv technical report).
- DeepSeek-V3 generated 60 tokens per second, 3x faster than V2 (V3 release).
- DeepSeek-R1-Zero’s AIME 2024 score rose from 15.6% to 77.9% during reinforcement learning (Nature).
- With self-consistency decoding, R1-Zero reached 86.7% on AIME 2024 (Nature).
- DeepSeek open-sourced 6 small models distilled from R1, with the 32B and 70B versions on par with OpenAI-o1-mini (R1 release).
- DeepSeek-R1 is MIT licensed, and API outputs may be used for fine-tuning and distillation (R1 release).
DeepSeek model stats
- V3.1 added 840B tokens of continued pretraining to extend V3’s context to 128K (V3.1 release).
- V3.2 was trained on synthesized agent data covering 1,800+ environments and 85k+ complex instructions (V3.2 release).
- V3.2-Speciale reached gold-level results at IMO, CMO, the ICPC World Finals and IOI 2025 (V3.2 release).
- DeepSeek-V4-Pro has 1.6T total and 49B active parameters (V4 preview).
- DeepSeek-V4-Flash has 284B total and 13B active parameters (V4 preview).
- 1M context became the default across all official DeepSeek services with V4 (V4 preview).
- DeepSeek-V4.1-Flash is a 552B-parameter MoE with 8B active parameters for input and 16B for output (V4.1-Flash release).
- V4.1-Flash’s KV cache needs 1/4 the HBM and 1/8 the SSD storage of the previous generation (V4.1-Flash release).
- Since September 14, 2026, all deepseek-v4-pro requests route to V4.1-Flash at Flash rates (V4.1-Flash release).
- The legacy deepseek-chat and deepseek-reasoner API names retired after July 24, 2026 (V4 preview).
- DeepSeek-V2.5 raised its internal safety score to 82.6% from 74.4% and cut safety spillover to 4.6% from 11.3% (V2.5 release).
- DeepSeek function calling supports up to 128 functions in one call (API upgrade).
DeepSeek API pricing stats
- deepseek-flash costs $0.15 per million cache-miss input tokens off-peak and $0.3 at peak (pricing).
- deepseek-flash output costs $0.6 per million tokens off-peak and $1.2 at peak (pricing).
- deepseek-v4-pro costs $0.66 per million cache-miss input tokens off-peak and $1.32 at peak, with output at $1.98 and $3.96 (pricing).
- Cache-hit input on deepseek-flash costs $0.003 per million tokens off-peak (pricing).
- Off-peak rates are half of peak rates (pricing); DeepSeek introduced the split with the V4 lineup, off-peak 50% lower (V4-Pro GA).
- Both current models offer 1M context and up to 384K output tokens (pricing).
- Concurrency limits are 2500 for deepseek-flash and 500 for deepseek-v4-pro (pricing).
- DeepSeek’s Search and Fetch APIs cost $0.75 per 1,000 successful requests each (pricing).
- Search and Fetch calls through DeepSeek Harness, Claude Code or Codex are free until per-request billing starts on November 1, 2026, Beijing time (pricing).
- DeepSeek-R1 launched at $0.55 per million input tokens (cache miss), $0.14 on cache hits and $2.19 per million output tokens (R1 release).
- From February 8, 2025, V3 cost $0.27 per million input tokens, $0.07 on cache hits and $1.10 for output (V3 release).
- DeepSeek cut API prices by 50%+ with V3.2-Exp in September 2025 (V3.2-Exp release).
- Context caching on disk cut cache-hit input to $0.014 per million tokens, up to 90% cheaper (context caching).
- Users saved over 50% on average with context caching, even without optimizing for it (context caching).
- First-token latency on a 128K prompt fell from 13s to 500ms with caching (context caching).
- The Search API runs on DeepSeek’s own search engine, the same one serving its chat app and agents (V4.1-Flash-1017 release).
| Model | Input, cache miss | Output |
|---|---|---|
| deepseek-flash off-peak | $0.15 | $0.6 |
| deepseek-flash peak | $0.3 | $1.2 |
| deepseek-v4-pro off-peak | $0.66 | $1.98 |
| deepseek-v4-pro peak | $1.32 | $3.96 |
Per million tokens. Source: DeepSeek Models & Pricing, checked October 26, 2026.
What to do with this data
- Don’t confuse downloads with buyers. Fifty million R1 downloads are developers and companies running the model themselves. They tell you DeepSeek matters in the model supply chain, not that your buyers ask DeepSeek which product to buy.
- Watch DeepSeek’s search engine. DeepSeek now runs its own web index and sells Search and Fetch to agents built on Claude Code and Codex, so pages that are crawlable and clearly structured get picked up by more AI agents. Our AI crawler statistics cover the bots doing that work.
- Measure the engines that send pipeline first. Mentionova tracks ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews and Reddit, and its agents draft the briefs, content and Reddit replies that close citation gaps. It is #1 in our best AI visibility tools and AI visibility tools for agencies roundups; see Mentionova pricing for plans from $99 a month.
- Compare the engines. Read these numbers next to our ChatGPT statistics, Claude statistics, Gemini statistics, Grok statistics and AI chatbot statistics.
Frequently asked questions
How many people use DeepSeek?
DeepSeek does not publish user counts. Its search engine serves hundreds of millions of requests a day, DeepSeek-R1 has 50,565,681 all-time downloads on Hugging Face, and NIST found DeepSeek model downloads grew nearly 1,000% after January 2025.
How much did DeepSeek cost to train?
DeepSeek’s technical report says V3’s full training took 2.788M H800 GPU hours, which it prices at $5.576M assuming $2 per GPU hour. That figure covers the final training run, not research, staff or hardware purchases.
How much does the DeepSeek API cost?
deepseek-flash costs $0.15 per million cache-miss input tokens and $0.6 per million output tokens off-peak, double that at peak. deepseek-v4-pro costs $0.66 and $1.98 off-peak. Search and Fetch cost $0.75 per 1,000 requests each.
Is DeepSeek free?
The DeepSeek app is free, with no ads and no in-app purchases, per DeepSeek’s January 2025 launch note. The API is pay as you go.
Is DeepSeek safe to use?
NIST’s CAISI found DeepSeek’s most secure model, R1-0528, answered 94% of overtly malicious requests under a common jailbreak, against 8% for U.S. reference models, and its agents were 12 times more likely to be hijacked.
How we compiled these statistics
We fetched 25 primary source pages on October 26, 2026 and recorded 60 data points in docs/content/stats-ledgers/deepseek-statistics.json. Every numbered figure above maps to a ledger row with a verbatim quote. Model, training, pricing and search figures are DeepSeek’s own statements from its API docs news posts, its pricing page, its V3 technical report on arXiv and its R1 paper in Nature. Download, like, star and follower counts come straight from the Hugging Face and GitHub APIs on the fetch day and will keep rising. The safety and performance comparisons are from NIST’s Center for AI Standards and Innovation. Benchmark claims in DeepSeek’s release notes are self-reported. We did not use market-research sites or traffic estimators. Related reading: AI search statistics, Perplexity statistics, Microsoft Copilot statistics.