DeepSeek AI Statistics 2026: Users, Benchmarks and Funding

DeepSeek had 130 million monthly active users in China, behind Doubao at 382 million and Qwen at 167 million, according to QuestMobile’s 2026 first-half report. Those DeepSeek AI statistics sit beside a capability gap: NIST’s Center for AI Standards and Innovation found that DeepSeek V4’s capabilities lag behind the frontier by about 8 months, even as DeepSeek keeps shipping cheaper models.

The newest of those models, DeepSeek-V4.1-Flash, is a 552 billion-parameter MoE with just 8 billion active parameters for input and 16 billion for output. The figures below cover users, downloads, market share, pricing, benchmarks, security testing, funding, and government actions, with self-reported and independent numbers side by side.

Key Takeaways

• DeepSeek’s own data puts DeepSeek V4 about as capable as Opus 4.6 and GPT-5.4, while CAISI’s evaluations indicate that DeepSeek V4 performs similarly to GPT-5, which was released about 8 months ago.
• DeepSeek holds 89% market share in China, 56% in Belarus, 49% in Cuba, and 43% in Russia in Microsoft’s AI diffusion data.
• On Hugging Face, DeepSeek-R1 has 48,931,630 all-time downloads, more than any other DeepSeek checkpoint tracked here.
• DeepSeek’s most secure model (R1-0528) responded to 94% of overtly malicious requests when a common jailbreaking technique was used, compared with 8% of requests for U.S. reference models.
• DeepSeek-V4.1-Flash output tokens cost $1.2 per 1 million at peak and $0.6 off-peak.
• DeepSeek is raising funds at a valuation close to 500 billion yuan ($74 billion), people familiar with the matter told Bloomberg.

Editor’s Choice

• China monthly active users: 130 million for DeepSeek in QuestMobile’s 2026 first-half report.
• DeepSeek-V4-Pro size: 1.6 trillion total and 49 billion active parameters.
• First-round valuation: more than 350 billion yuan ($52 billion) post-money.
• Default context window: 1 million tokens across all official DeepSeek services.
• V4.1-Flash training corpus: a multimodal corpus comprising 45 trillion tokens for DeepSeek-V4.1-Flash.
• DSec sandbox throughput: over 5,000 sandbox creations per second and over 380,000 concurrent sandboxes.
• Model download growth: downloads of DeepSeek models on model-sharing platforms have increased nearly 1,000% since January 2025.

DeepSeek AI Statistics: Model Lineup and Parameter Counts

These figures are compiled from 29 cited findings: 19 from DeepSeek documentation, the Hugging Face API, NIST, government and regulator pages, and Microsoft (tier 1), 7 from arXiv, Nature, and a16z (tier 2), and 3 from news coverage (tier 3). Sources run from December 2024 to September 2026. Only primary documents or named measurement data qualified, and figures are reviewed on a rolling basis when sources publish new editions.

• Both V4 models support 1 million-token context and dual thinking and non-thinking modes.
• The legacy deepseek-chat and deepseek-reasoner models were set to be fully retired and inaccessible after July 24, 2026.
• Compared with the previous generation, V4.1-Flash’s KV cache needs just 1/4 the HBM and 1/8 the SSD storage.
• DeepSeek-V4.1-Flash reduces its global KV cache footprint to 890 bytes per token.
• DeepSeek-V4.1-Flash and DeepSeek-V4-Pro-0813 both support a maximum output of 384,000 tokens.

Source: DeepSeek API Docs (2026), DeepSeek-V3 Technical Report (arXiv, 2024). V4.1-Flash active parameters shown for output.

DeepSeek Model Downloads on Hugging Face

• On Hugging Face, DeepSeek-R1 has 14,289 likes, against 5,593 for DeepSeek-V4-Pro and 4,251 for DeepSeek-V3.
• DeepSeek-V4.1-Flash has 3,747 likes within weeks of release.
• The five DeepSeek checkpoints below total about 93.5 million all-time downloads.
• DeepSeek-R1 has about 2.3 times the all-time downloads of DeepSeek-V3.
• CAISI found that the release of DeepSeek R1 has driven adoption of PRC models across the AI ecosystem.

Don’t chase tech news. We track it for you.

One weekly briefing with the launches, AI developments, and breaches that matter. No filler.

Email address

Subscribe Now

Read by pros at Fortinet, TSMC, Barclays, and Deloitte.

Recent Developments

• September 2026: DeepSeek released the DeepSeek-V4.1-Flash model and decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026.
• September 2026: DeepSeek presented DeepSeek Elastic Compute (DSec), a production sandbox platform for large-scale agentic training.
• August 2026: DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model, became available on the DeepSeek API platform.
• August 2026: The GA release of DeepSeek-V4-Pro rolled out on the APP, Web, and API with a Terminal Bench 2.1 score of 87.9.
• August 2026: Peak and off-peak API pricing took effect at 16:00 UTC on Aug 16, 2026, with off-peak rates 50% lower than peak.
• July 2026: DeepSeek raised about $7.4 billion in its first external funding round, according to a regulatory filing.

CAISI Independent Evaluation of DeepSeek V4

• DeepSeek V4 is the most capable PRC AI model evaluated by CAISI to date.
• Compared to GPT-5.4 mini, DeepSeek V4 was more cost-efficient on 5 out of 7 benchmarks.
• On the 7 benchmarks, DeepSeek V4 ranged from 53% less expensive to 41% more expensive.
• CAISI used 16 benchmarks across 35 models to fit its capability comparison.
• DeepSeek V4 Pro posted an IRT-estimated Elo of 800 (± 28), against 1,260 (± 28) for GPT-5.5.
• CAISI imputed one DeepSeek V4 Pro cyber score from a subset of samples via IRT.

Source: NIST CAISI Evaluation of DeepSeek V4 Pro, May 2026

Worth noting: DeepSeek V4 scores better on DeepSeek’s self-reported evaluations than on CAISI evaluations, which include non-public benchmarks. The widest gaps sit on held-out tests such as PortBench and the ARC-AGI-2 semi-private set, so vendor benchmark tables should be read as a ceiling, not a forecast of production results.

Is DeepSeek AI better than ChatGPT?

Not on CAISI’s tests of the underlying models. OpenAI GPT-5.5 scored 81% on SWE-Bench Verified and 71% on CTF-Archive-Diamond, against 74% and 32% for DeepSeek V4 Pro. DeepSeek’s stronger argument is cost, and the Claude vs ChatGPT data shows how the leading U.S. assistants compare.

DeepSeek Monthly Active Users

• China’s AI-native apps recorded 499 million monthly active users by May 2026, up 85.4% from a year earlier.
• Users of AI-native apps averaged 92.7 uses per month and 183 minutes of monthly usage time.
• DeepSeek’s user count equals about 26% of the 499 million AI-native app MAU total in China.
• On mobile, DeepSeek fell off its peak by 22%, a16z reported in its August 2025 ranking.
• On web, DeepSeek saw a drop-off of more than 40% from its peak in February 2025.

DeepSeek Market Share by Country

• DeepSeek’s web traffic splits across China (33.5%), Russia (7.1%), and the US (6.6%), according to a16z’s consumer AI ranking.
• DeepSeek is the only product that bridges the divide between markets served by U.S. assistants and China and Russia, according to a16z.
• Russia has emerged as a third pole with the second-highest rate of DeepSeek saturation.
• Countries where entrenched alternatives already meet local needs, such as Israel and South Korea, show minimal uptake.
• The absence of subscription fees or payment requirements lowered the barrier for millions of users, especially in price-sensitive regions.

DeepSeek V4.1-Flash Benchmark Scores

• DeepSeek-V4.1-Flash posted a Codeforces rating of 3471.
• It scored 36.8 on HLE and 63.9 on HLE with tools.
• It scored 30.0 on Terminal-Bench 3.0 and 31.2 on Terminal-Bench 4.0.
• It scored 20.3 on ProgramBench and 15.3 on ExploitGym.
• DeepSeek says new pre-training methods and larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro.

How accurate is DeepSeek?

Accuracy depends heavily on the test. In CAISI’s run, DeepSeek V4 Pro solved 90% of GPQA-Diamond tasks and 97% on OTIS-AIME-2025, but 44% on PortBench. Math and science questions suit DeepSeek best, while held-out software and cyber tasks expose the largest weaknesses.

DeepSeek Agent Benchmark Progression

• DeepSeek-V4-Flash launched with 284 billion total and 13 billion active parameters.
• DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.
• The V4-Flash update scored 54.2 on NL2Repo and 54.4 on DeepSWE.
• The V4-Pro GA version scored 61.5 on NL2Repo, 83.3 on Cybergym, and 74.1 on Toolathlon-Verified.
• DeepSeek-V4-Flash-Vision-Exp brings its multimodal agent capabilities close to Opus-4.8.
• V4.1-Flash’s Terminal-Bench 2.1 score sits 7.9 points above V4-Flash-0731.

Terminal Bench and NL2Repo measure multi-step tool use, the workload behind wider AI agent adoption.

DeepSeek API Pricing

• Off-peak rates are half of the peak rates, and peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday.
• deepseek-flash carries a concurrency limit of 2,500 versus 500 for deepseek-v4-pro.
• New V4.1-Flash pricing took effect at 04:00 UTC on Sept 10, 2026.
• DeepSeek-V4-Pro output tokens cost 3.3 times the V4.1-Flash rate at peak.
• The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price.

Source: DeepSeek Models and Pricing, September 2026

Why it matters: CAISI’s 2025 evaluation found that one U.S. reference model costs 35% less on average than the best DeepSeek model to perform at a similar level. The V4 generation reversed that result in CAISI’s own retest, which is why price, not raw capability, now anchors DeepSeek’s pitch to developers.

DeepSeek Security and Jailbreak Findings

• Hijacked agents sent phishing emails, downloaded and ran malware, and exfiltrated user login credentials, all in a simulated environment.
• In software engineering and cyber tasks, the best U.S. model evaluated solves over 20% more tasks than the best DeepSeek model.
• The House Select Committee on the CCP reported that DeepSeek funnels Americans’ data to the PRC through backend infrastructure connected to a U.S. government-designated Chinese military company.
• The Committee also found that DeepSeek covertly manipulates the results it presents to align with CCP propaganda.

Source: NIST CAISI Evaluation of DeepSeek AI Models, September 2025

Those jailbreak results are worth reading alongside the wider AI jailbreak data for other model families before deploying any open-weight model in an agent that can send email or run code.

DeepSeek-R1 Reasoning Results

• The reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories, the DeepSeek-R1 authors reported.
• The DeepSeek-R1 paper appeared in Nature volume 645, pages 633-638 (2025).
• The R1-Zero result greatly surpasses the average performance across all human competitors of the AIME.
• DeepSeek-R1-Zero also achieves remarkable performance in coding competitions and graduate-level biology, physics, and chemistry problems.
• DeepSeek invited independent researchers in September 2025 to rigorously test the model’s performance, aiming for publication in Nature.

DeepSeek Training Data and Compute

• DeepSeek-V3 has 671 billion total parameters with 37 billion activated for each token and was pre-trained on 14.8 trillion tokens.
• DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching.
• DeepSeek invites partners planning a large-scale deployment with 2,000 GPUs plus a storage cluster.
• The V4.1-Flash pre-training corpus is about 3 times the size of the V3 corpus, though V4.1-Flash’s corpus is multimodal and V3’s was text.

Source: DeepSeek-V3 Technical Report (2024) and DeepSeek-V4.1-Flash paper (2026), arXiv

DeepSeek Agent Training Infrastructure

• A single production-scale unit of DSec spans around 160 nodes.
• DSec exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK.
• DSec loads image data on demand from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem.
• DSec is co-designed with the reinforcement learning (RL) framework and mitigates agent misbehavior such as reward hacking.
• DSec decouples stateful rollout execution from preemptible GPU training.

Source: DeepSeek Elastic Compute (DSec) paper, arXiv, September 2026

DeepSeek Funding and Valuation

• DeepSeek raised about $7.4 billion in its first external funding round.
• The details emerged in a disclosure by Anhui Korrun Co. Ltd. (300577.SZ), a Chinese luggage and travel-products maker whose wholly owned subsidiary invested indirectly in DeepSeek through a private-equity fund.
• The Hangzhou-based startup paused the process following frustration about leaked remarks from the company’s founder to investors, before reopening it in August.
• Our tracker of recent AI funding rounds shows where DeepSeek’s round sits against U.S. and Chinese peers as new deals close.

Source: Anhui Korrun filing via Caixin (July 2026), Bloomberg (August 2026)

Government Actions on DeepSeek

• PSPF Direction 001-2025, published 4 February 2025, requires Australian Government entities to prevent the use or installation of DeepSeek products, applications and web services.
• The Berlin Commissioner for Data Protection and Freedom of Information notified Google and Apple in Germany of the AI app DeepSeek as illegal content, citing the unlawful transfer of personal data from users of the app to China.
• Hangzhou DeepSeek Artificial Intelligence Co., Ltd. has no establishment in the European Union.
• The House Select Committee reported that it is highly likely that DeepSeek used unlawful model distillation techniques to create its model.
• DeepSeek’s AI model reportedly utilizes tens of thousands of chips that are currently restricted from export to the PRC, the Committee said.

Device bans run alongside wider AI regulation by jurisdiction, which sets the terms on which any foreign model can be offered commercially.

Conclusion

DeepSeek counted 130 million monthly active users in China, and its open-weight checkpoints keep spreading well beyond that base. Yet CAISI evaluations indicate that DeepSeek V4’s capabilities lag behind the frontier by about 8 months. The gap between DeepSeek’s reach and its independently measured capability is the defining pattern in these statistics, and it matters most for developers choosing a model for agentic or security-sensitive work.

DeepSeek’s V4.1-Flash notice points to a V4.1-Pro launch still to come, so the next flagship release, and whether CAISI retests it, is the figure to watch among recent model releases.