Links marked with * are affiliate links. If a purchase is made through such links, we receive a commission.
The often-quoted $5.576 million tells only part of the story.
That's the math for the 2.788 million GPU hours behind V3's training, at an assumed $2 per hour. Earlier research and ablation experiments aren't included.
If you're looking up DeepSeek's revenue in 2026, the numbers look very different. According to a media report, DeepSeek hit an annualized revenue run rate of $1 billion in September 2026.
Here are the key DeepSeek statistics for 2026, with numbers and facts on users, models, API pricing, and funding. The short version first:
TL;DRKey Takeaways
QuestMobile put DeepSeek's active-user scale in China's native AI apps at 123 million in September 2026, down from 129 million in June. Sensor Tower reported 127 million worldwide mobile app downloads by the end of June 2025.
DeepSeek's V3 report lists 2.788 million H800 GPU hours. At an assumed $2 per GPU hour, that works out to $5.576 million. Earlier research and ablation experiments are excluded.
On output tokens, DeepSeek sits well below the flagships of OpenAI, Anthropic, Google, and xAI. Rates differ for cache hits, uncached input, and peak versus off-peak hours.
1. Company and origins
According to an Associated Press report, DeepSeek was founded in Hangzhou in 2023. Its CEO, Liang Wenfeng, had previously co-founded the quantitative hedge fund High-Flyer.
2. Downloads and user numbers
DeepSeek's reach is documented by Sensor Tower download estimates and QuestMobile user figures for China.
Sensor Tower also reported that DeepSeek reached number one in the US App Store on January 27, 2025, and in Google Play in the US the following day. For the first 19 days, its February 2025 report gives the download shares shown in the chart below.
Download share in four named countries, first 19 days after launch (January 2025)
The figures below are historical numbers from DeepSeek's own reports.
3.1 DeepSeek-V3: training and original comparison
The DeepSeek-V3 technical report lists 671 billion parameters in total and 37 billion activated per token. Pretraining used 14.8 trillion tokens.
The report records 2.788 million H800 GPU hours across pretraining, context extension, and post-training. At an assumed rental price of $2 per GPU hour, that works out to $5.576 million. The report excludes earlier research and ablation experiments.
The chart reproduces selected results from the original DeepSeek-V3 study. It compares the results with GPT-4o and Claude 3.5 Sonnet.
DeepSeek-V3 compared with GPT-4o and Claude 3.5 Sonnet
For the DeepSeek-V3 base model, the same report separately lists 87.5% on BBH (exact match, 3-shot).
3.2 Reasoning models
DeepSeek’s R1 model card reports 90.8% on MMLU (Pass@1) and 79.8% on AIME 2024 (Pass@1). The R1-0528 release note compares the earlier R1 at 70.0% with R1-0528 at 87.5% on AIME 2025.
DeepSeek-V3.2, December 2025; V4 reports no AIME 2025 value
The chart adds V3.2 from DeepSeek's technical report to the results in the May 28, 2025 release note. On AIME 2025, the token budgets differ there, with 23,000 generated tokens for V3.2-Speciale and 13,000 generated tokens for GPT-5 High.
3.3 DeepSeek-V4 and V4.1 Flash
DeepSeek previewed V4 on April 24, 2026. The official V4 report describes V4 Pro with 1.6 trillion parameters, 49 billion of them active. It lists V4 Flash at 284 billion parameters, with 13 billion active. Both versions support a context window of up to 1 million tokens, according to the report.
In max mode, DeepSeek's V4 report gives V4 Pro 87.5% on MMLU-Pro and 90.1% on GPQA Diamond. The provider gave a 93.5% LiveCodeBench result.
The model cards list each model's total parameters and the parameters activated per token. For these Mixture-of-Experts models, the two figures differ by a factor of roughly 18 to 33. V4.1 Flash is left out because its model card gives 552 billion parameters for the backbone only, plus 196 billion for Engram memory.
Total
Active per token
billion parameters
DeepSeek-V3
Total: 671 B
Active per token: 37 B
DeepSeek-R1
Total: 671 B
Active per token: 37 B
DeepSeek-V4 Flash
Total: 284 B
Active per token: 13 B
DeepSeek-V4 Pro
Total: 1.6 T
Active per token: 49 B
Source: Hugging Face model cards
|
CC BY 4.0
gradually.ai
DeepSeek-V3, Model card, retrieved October 27, 2026
DeepSeek-R1, Model card, retrieved October 27, 2026
V4 Pro became generally available on August 13, 2026. DeepSeek released V4.1 Flash on September 10. Its model card describes a multimodal Mixture-of-Experts model with image processing and a context window of up to 1 million tokens. It activates 8 billion at prefill and 16 billion at decode.
According to the September 10 release note, V4 Flash and V4 Vision Experimental were retired. Legacy API calls under the V4 Flash name are routed to V4.1 Flash. The table below compares the September 10 model card with the card for V4.1 Flash-1017. The Codeforces rating is 3,471 rating points.
According to the official V4.1 Flash model card, DeepSeek ran most tasks at maximum reasoning effort. The Flash-1017 card restates only three of the values; “n/a” means no value is listed there.
Note
The model card calls 552 billion the backbone and lists another 196 billion parameters for Engram memory. The Hugging Face page also displays 763 billion parameters as the model size.
4. API pricing
The official pricing page lists V4.1 Flash-1017 under API model ID “deepseek-flash” and V4 Pro as “deepseek-v4-pro”. The charts show the rates listed there on October 17, 2026, in US dollars per one million tokens, split by cache hits, uncached input, output, and peak versus off-peak.
DeepSeek defines one weekday peak window from 01:00 to 04:00 UTC. A second runs from 06:00 to 10:00 UTC. Weekends and Chinese public holidays use off-peak rates, according to the pricing page.
4.1 V4.1 Flash-1017 rates
These are the V4.1 Flash-1017 rates in US dollars per one million tokens, off-peak and peak:
The September 10 launch note initially said that requests using “deepseek-v4-pro” would route to V4.1 Flash at Flash rates from September 14 until V4.1 Pro launches.
A later entry in the official changelog says DeepSeek would continue Pro API service after September 14 at users’ request, with billing unchanged. The current pricing page also lists V4 Pro with its own model version and rates.
4.3 Output prices compared with other providers
On output tokens, DeepSeek sits well below the flagships of OpenAI, Anthropic, Google, and xAI. The chart shows each provider's current flagship and a cheaper tier at the base rate up to the first context threshold. For DeepSeek, the off-peak rate applies.
Audited annual accounts are not available. The following business figures come from media reports.
Revenue
Reported figure
475 million yuan
Period and qualification
January through July 2026, attributed by The Information to two informed sources
Annualized revenue run rate
Reported figure
$1 billion, according to a media report
Period and qualification
Reported in September 2026 by The Information citing two sources; Reuters could not independently verify it; not directly comparable to the 475 million yuan for January through July
Net loss
Reported figure
715 million yuan
Period and qualification
January through July 2026, same media report
Gross margin
Reported figure
44.6% overall; 82.9% for API business
Period and qualification
January through July 2026, same media report
First external funding round
Reported figure
About 50 billion yuan (about $7.4 billion)
Period and qualification
June 2026; Reuters reported that payments had been made by June 17
Transaction-implied value from a filing
Reported figure
350.88 billion yuan; not a definitive round valuation
Period and qualification
Reuters analysis of an investor filing dated July 16, 2026; Reuters reported an estimate of about 450 billion yuan from sources the day before
Further funding round
Reported figure
About 100 billion yuan (about $15 billion), reported, not closed
Period and qualification
Reuters report dated October 22, 2026, citing four people with knowledge; figures not final; on October 6 Reuters had cited more than 80 billion yuan
Item
Reported figure
Period and qualification
Revenue
475 million yuan
January through July 2026, attributed by The Information to two informed sources
Annualized revenue run rate
$1 billion, according to a media report
Reported in September 2026 by The Information citing two sources; Reuters could not independently verify it; not directly comparable to the 475 million yuan for January through July
Net loss
715 million yuan
January through July 2026, same media report
Gross margin
44.6% overall; 82.9% for API business
January through July 2026, same media report
First external funding round
About 50 billion yuan (about $7.4 billion)
June 2026; Reuters reported that payments had been made by June 17
Transaction-implied value from a filing
350.88 billion yuan; not a definitive round valuation
Reuters analysis of an investor filing dated July 16, 2026; Reuters reported an estimate of about 450 billion yuan from sources the day before
Further funding round
About 100 billion yuan (about $15 billion), reported, not closed
Reuters report dated October 22, 2026, citing four people with knowledge; figures not final; on October 6 Reuters had cited more than 80 billion yuan
On September 24, Reuters reported that The Information put DeepSeek’s annualized revenue run rate at $1 billion. A run rate extrapolates current revenue across a year. The Information attributed some growth to price increases of 2.3 to 4.5 times.
On October 6, 2026, Reuters reported, citing one person with knowledge of the round and following an earlier Bloomberg report, that commitments exceeded 80 billion yuan, above an original 50 billion yuan target. Reuters gave a target valuation of 500 billion yuan, or about $74 billion, and said a close was expected in October.
On October 22, 2026, Reuters reported, citing four people with knowledge, that DeepSeek had raised about 100 billion yuan (about $15 billion) in the round. CATL and Tencent, both existing shareholders, were among the biggest investors; GigaDevice, Meituan and Geely were also named. The fundraising was set to complete in the coming days, but the details were not final and could change. Reuters could not immediately verify the report, and DeepSeek and the investors did not comment. Dow Jones reported the same day that DeepSeek had raised $15 billion; DeepSeek had not confirmed this as of October 31, 2026. On an IPO, Reuters wrote that the company aimed to kick off the process this year.
On January 27, 2025, NVIDIA shares fell 17% in one trading day, according to the Associated Press. Reuters reported at the time that the market-capitalization loss was on track to reach about $600 billion. Its report connected the move with investor concerns about DeepSeek’s reported model efficiency.
The chart shows how closing prices on January 27, 2025, changed from the close on January 24. The range runs from losses of about 17% at Broadcom and NVIDIA to gains at Apple and Meta.
DeepSeek says in its privacy policy that it directly collects, processes, and stores personal information in the People’s Republic of China.
The official V3 repository licenses its code under MIT, but its README says use of the V3 Base and Chat models is subject to a separate Model License. The official model repositories for R1, V4 Pro, and V4.1 Flash specify the MIT license for the model weights.
When weights run on your own device without an external inference service, prompts do not need to be sent to DeepSeek. Data flows still depend on the local runtime, logging, and connected services.
8. DeepSeek timeline
The timeline runs from the company's founding in 2023 to V4.1 Flash-1017 in October 2026.
2023
DeepSeek is founded in Hangzhou
December 26, 2024
DeepSeek-V3 is introduced
January 15, 2025
DeepSeek formally announces its app
January 20, 2025
DeepSeek-R1 is introduced
May 28, 2025
DeepSeek-R1-0528 is released
September 29, 2025
DeepSeek-V3.2-Exp is released as an experimental version
DeepSeek scores 85 out of 100 in the AI vendor popularity index, rank 4 of 39. The value combines search demand for the vendor and its AI brands, usage via OpenRouter and Hugging Face and mentions in tech media.
In the EU, DeepSeek is at rank 5 of 36, in Germany at rank 5 of 36. Use the switch above the chart to change between world, EU and Germany.
Vendor popularity index
0 to 100, higher is more popular
OpenAI97
Google91
Anthropic88
DeepSeek85
Meta79
Alibaba78
Perplexity75
ElevenLabs73
MiniMax72
Z.ai72
Moonshot AI72
Microsoft71
Mistral AI70
xAI70
Stability AI69
ByteDance68
NVIDIA68
Tencent66
Black Forest Labs64
Cohere62
Liquid AI61
Skywork AI61
Kuaishou60
Xiaomi60
Amazon60
Runway60
StepFun60
IBM57
ShengShu Technology57
Mapika57
Baidu56
inclusionAI53
Upstage49
Bilibili Index Team48
Nace.AI47
AI2147
Arcee AI47
Institute of Foundation Models45
Nex AGI37
020406080100
OpenAI, OpenAI: 9797
Google, Google: 9191
Anthropic, Anthropic: 8888
DeepSeek, DeepSeek: 8585
Meta, Meta: 7979
Alibaba, Alibaba: 7878
Perplexity, Perplexity: 7575
ElevenLabs, ElevenLabs: 7373
MiniMax, MiniMax: 7272
Z.ai, Z.ai: 7272
Moonshot AI, Moonshot AI: 7272
Microsoft, Microsoft AI: 7171
Mistral AI, Mistral AI: 7070
xAI, xAI: 7070
Stability AI, Stability AI: 6969
ByteDance, ByteDance: 6868
NVIDIA, NVIDIA: 6868
Tencent, Tencent: 6666
Black Forest Labs, Black Forest Labs: 6464
Cohere, Cohere: 6262
Liquid AI, Liquid AI: 6161
Skywork AI, Skywork AI: 6161
Kuaishou, Kuaishou: 6060
Xiaomi, Xiaomi: 6060
Amazon, Amazon: 6060
Runway, Runway: 6060
StepFun, StepFun: 6060
IBM, IBM: 5757
ShengShu Technology, ShengShu Technology: 5757
Mapika, Mapika: 5757
Baidu, Baidu: 5656
inclusionAI, inclusionAI: 5353
Upstage, Upstage: 4949
Bilibili Index Team, Bilibili Index Team: 4848
Nace.AI, Nace.AI: 4747
AI21, AI21: 4747
Arcee AI, Arcee AI: 4747
Institute of Foundation Models, Institute of Foundation Models: 4545
Nex AGI, Nex AGI: 3737
OpenAI
Google
Anthropic
DeepSeek
Meta
Alibaba
Perplexity
ElevenLabs
MiniMax
Z.ai
Moonshot AI
Microsoft
Mistral AI
xAI
Stability AI
ByteDance
NVIDIA
Tencent
Black Forest Labs
Cohere
Liquid AI
Skywork AI
Kuaishou
Xiaomi
Amazon
Runway
StepFun
IBM
ShengShu Technology
Mapika
Baidu
inclusionAI
Upstage
Bilibili Index Team
Nace.AI
AI21
Arcee AI
Institute of Foundation M…
Nex AGI
As of Nov 21, 2026 · Source: gradually.ai popularity index
10. How popular are DeepSeek's models?
DeepSeek's most popular model is DeepSeek-V4.1-Flash with 85 out of 100 (rank 10 of 881 AI models with a popularity score). In total, 24 DeepSeek models have a score; the chart highlights them.
In the EU, DeepSeek-V4.1-Flash is at rank 9 of 613, in Germany, DeepSeek-V4.1-Flash is at rank 8 of 692. Use the switch above the chart to change between world, EU and Germany.
Popularity of DeepSeek models
Popularity index from 0 to 100, higher is more popular.
28 of 117 models
Claude Opus 5.588
GPT-6.1 Terra87
GLM-5.387
Kimi K386
Nano Banana 286
Claude Haiku 5.586
GPT-6 Luna86
Gemini 4 Argon85
GPT-6.1 Sol85
DeepSeek-V4.1-Flash85
GLM-5.3-Flash84
Sora 284
GPT-6 Astra84
MiniMax H384
Nano Banana Pro82
Veo 3.182
Qwen 3.8 27B82
GPT-581
DeepSeek V4.1 Flash 101781
Gemini 4 Carbon81
DeepSeek-V4-Flash80
DeepSeek-V4-Pro74
DeepSeek-R172
DeepSeek V4 Flash 073168
DeepSeek-V368
DeepSeek-V3.263
DeepSeek-R1-Distill-Qwen-32B53
DeepSeek OCR 251
020406080100
Claude Opus 5.5, Anthropic: 8888
GPT-6.1 Terra, OpenAI: 8787
GLM-5.3, Z.ai: 8787
Kimi K3, Moonshot AI: 8686
Nano Banana 2, Google: 8686
Claude Haiku 5.5, Anthropic: 8686
GPT-6 Luna, OpenAI: 8686
Gemini 4 Argon, Google DeepMind: 8585
GPT-6.1 Sol, OpenAI: 8585
DeepSeek-V4.1-Flash, DeepSeek: 8585
GLM-5.3-Flash, Z.ai: 8484
Sora 2, OpenAI: 8484
GPT-6 Astra, OpenAI: 8484
MiniMax H3, MiniMax: 8484
Nano Banana Pro, Google: 8282
Veo 3.1, Google: 8282
Qwen 3.8 27B, Alibaba: 8282
GPT-5, OpenAI: 8181
DeepSeek V4.1 Flash 1017, DeepSeek: 8181
Gemini 4 Carbon, Google DeepMind: 8181
DeepSeek-V4-Flash, DeepSeek: 8080
DeepSeek-V4-Pro, DeepSeek: 7474
DeepSeek-R1, DeepSeek: 7272
DeepSeek V4 Flash 0731, DeepSeek: 6868
DeepSeek-V3, DeepSeek: 6868
DeepSeek-V3.2, DeepSeek: 6363
DeepSeek-R1-Distill-Qwen-32B, DeepSeek: 5353
DeepSeek OCR 2, DeepSeek: 5151
Claude Opus 5.5
GPT-6.1 Terra
GLM-5.3
Kimi K3
Nano Banana 2
Claude Haiku 5.5
GPT-6 Luna
Gemini 4 Argon
GPT-6.1 Sol
DeepSeek-V4.1-Flash
GLM-5.3-Flash
Sora 2
GPT-6 Astra
MiniMax H3
Nano Banana Pro
Veo 3.1
Qwen 3.8 27B
GPT-5
DeepSeek V4.1 Flash 1017
Gemini 4 Carbon
DeepSeek-V4-Flash
DeepSeek-V4-Pro
DeepSeek-R1
DeepSeek V4 Flash 0731
DeepSeek-V3
DeepSeek-V3.2
DeepSeek-R1-Distill-Qwen-…
DeepSeek OCR 2
Source: gradually.ai popularity index
11. How good are DeepSeek's models?
In the gradually.ai overall index, which combines independent benchmarks into one value, DeepSeek V4.1 Flash 1017 is DeepSeek's strongest model: rank 21 of 99 with 106.9 points. 4 DeepSeek models are ranked in the index. 100 points is the average of a fixed reference group; 15 points more means a model wins about three of four benchmark comparisons.
Artificial Analysis uses AA-Omniscience to measure how often a model gives a wrong answer to a knowledge question instead of admitting it does not know. DeepSeek's best result comes from DeepSeek V4.1 Flash 1017: 25% of its answers that are not correct are wrong guesses (rank 15 of 76, rank 1 hallucinates least). The median of all measured models is 40%.
2 DeepSeek models have a measurement. Each model counts with its best measured setting. The rate describes reliability, not safety in the sense of jailbreaks or harmful content.
Hallucination rate
Share of wrong answers among all answers that are not correct, lower is better. Best measured setting per model.
DeepSeek V4.1 Flash 1017 is the fastest with 137 output tokens per second. The median of all measured models is 70 tokens per second. OpenRouter has measured 6 DeepSeek models.
Speed
24 of 100 models
Inkling Small203tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Nemotron 3.5 Lightning161tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Claude Haiku 5.5160tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
MiniMax M2.7142tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Gemini 3.1 Flash-Lite140tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
GPT OSS 120B138tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Gemini 3.5 Flash-Lite138tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek V4.1 Flash 1017137tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek-V4.1-Flash137tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Gemini 3.5 Flash132tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Inkling118tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Gemini 3.6 Flash115tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Claude Sonnet 5.5109tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Ling 3.0 Flash108tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Mistral Small 4107tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
GPT OSS 20B104tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Grok 4.20 Reasoning104tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Mistral Medium 3.599tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
GPT-4.1 nano99tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
Gemini 3.1 Pro Preview99tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek-V4-Pro45tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek-V332tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek-V3.232tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
DeepSeek-V3 032429tokens/s · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21
tokens/s0100200300
Inkling Small, Thinking Machines Lab: 203 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21203
Nemotron 3.5 Lightning, NVIDIA: 161 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21161
Claude Haiku 5.5, Anthropic: 160 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21160
MiniMax M2.7, MiniMax: 142 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21142
Gemini 3.1 Flash-Lite, Google: 140 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21140
GPT OSS 120B, OpenAI: 138 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21138
Gemini 3.5 Flash-Lite, Google: 138 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21138
DeepSeek V4.1 Flash 1017, DeepSeek: 137 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21137
DeepSeek-V4.1-Flash, DeepSeek: 137 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21137
Gemini 3.5 Flash, Google: 132 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21132
Inkling, Thinking Machines Lab: 118 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21118
Gemini 3.6 Flash, Google: 115 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21115
Claude Sonnet 5.5, Anthropic: 109 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21109
Ling 3.0 Flash, inclusionAI: 108 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21108
Mistral Small 4, Mistral AI: 107 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21107
GPT OSS 20B, OpenAI: 104 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21104
Grok 4.20 Reasoning, xAI: 104 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-21104
Mistral Medium 3.5, Mistral AI: 99 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2199
GPT-4.1 nano, OpenAI: 99 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2199
Gemini 3.1 Pro Preview, Google: 99 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2199
DeepSeek-V4-Pro, DeepSeek: 45 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2145
DeepSeek-V3, DeepSeek: 32 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2132
DeepSeek-V3.2, DeepSeek: 32 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2132
DeepSeek-V3 0324, DeepSeek: 29 tokens per second · Median across providers on OpenRouter from the readings of the last 24 hours (each covering the last 30 minutes), excluding time to first token · Retrieved 2026-11-2129
Via the API, DeepSeek's cheapest model, DeepSeek-V4.1-Flash, costs $0.26 per 1M tokens. The most expensive is DeepSeek-R1 at $1.15. Both are blended prices of three parts input to one part output, as in the LLM price index.
API prices
Blended price in US dollars per 1M tokens, 3 parts input to 1 part output
DeepSeek is a Chinese AI company that develops language models and a chat service. The model families covered here include V3, R1, and the V4 series.
Hangzhou DeepSeek Artificial Intelligence Co., Ltd. operates the DeepSeek API service, according to its terms of use. The company describes its team in the official model documentation as a research group for foundational model technology; that documentation names no individual founders. According to the Associated Press, DeepSeek was founded in Hangzhou in 2023, and CEO Liang Wenfeng had previously co-founded the hedge fund High-Flyer.
V4.1 Flash-1017, which DeepSeek released on October 17, 2026 as an update to V4.1 Flash. For API calls, the provider recommends the alias `deepseek-flash`; the call and prices stayed the same, according to the provider. V4 Pro is listed separately as `deepseek-v4-pro`.
QuestMobile's Q3 report from October 20, 2026, puts DeepSeek at about 123 million active users in China's native AI app market for September. Its June 2026 figure was 129 million. Sensor Tower estimated 127 million worldwide mobile app downloads by the end of June 2025.
Off-peak V4.1 Flash-1017 costs $0.15 per million uncached input tokens and $0.6 per million output tokens, according to the official pricing page. V4 Pro is listed separately at $0.66 per million input tokens and $1.98 per million output tokens. Rates also depend on cache status and whether the request falls in the off-peak period. You can estimate a workload with the API cost calculator.
Partly. DeepSeek publishes downloadable weights for several model families. The official V3 repository licenses its code under MIT, but says use of the V3 Base and Chat models is subject to a separate Model License. Official sources list MIT licenses for R1, V4 Pro, and V4.1 Flash; R1 also notes inherited licenses for some Qwen- and Llama-derived models. This does not make the hosted chat service or all training data open source.
Yes, the DeepSeek app is listed in Germany's Apple App Store and on Google Play. DeepSeek's privacy policy says directly collected personal information may be processed and stored in China.
The figures come from provider reports and are not an independent ranking. DeepSeek reported 42.3 for V4.1 Flash-1017 on Terminal-Bench 4.0. Earlier V3, R1, and V3.2 scores are not measurements of the current model generation.
Finn Hillebrandt is the founder of Gradually AI, an SEO and AI expert. He helps online entrepreneurs simplify and automate their processes and marketing with AI. Finn shares his knowledge here on the blog in 50+ articles as well as through the AI Business Club.