DeepSeek is a Hangzhou-based AI lab, founded in July 2023 by hedge-fund billionaire Liang Wenfeng as a spin-off of his quant fund High-Flyer, that builds frontier open-weight language models at radically low cost.
Latest update Sep 30, 2026
01
Its January 2025 R1 reasoning model - released under an MIT license and trained for a small fraction of US-lab budgets - briefly made its free chatbot the top US iOS app and triggered a historic market rout in which Nvidia shed roughly $600B in a day. As of September 14, 2026 the live Flash model is DeepSeek-V4.1-Flash (552B MoE, 8B active on input / 16B on output, native vision; API name deepseek-flash), launched September 10, with V4-Pro API traffic routing to it from 04:00 UTC on September 14 until V4.1-Pro ships. V4-Pro (1.6T total / 49B active) had reached general availability on August 12, 2026, both MIT-licensed with 1M-token context and reportedly trained on Huawei Ascend chips. On August 21 it opened image inputs via DeepSeek-V4-Flash-Vision-Exp, and on August 31 it published those ~305B-parameter weights under an MIT license. DeepSeek serves ~130M monthly active users in China (QuestMobile, May 2026), third among domestic AI apps behind ByteDance's Doubao and Alibaba's Qwen, and monetizes through an API now billed on peak/off-peak rates, with V4.1-Flash off-peak at 50 percent of peak from September 10. After operating for three years on High-Flyer's money alone, it closed a record ~$7B maiden round at a ~$52B valuation in June 2026 - with most investors taking non-voting stakes that preserve Liang's control - and as of September 8, 2026 was still finalizing a follow-on raise of about 50B yuan (~$7.4B) at a ~$74B pre-money valuation that sources had expected to close by the end of August, while preparing a mainland IPO targeted for 2027.
deepseek.comUpdated Sep 14, 2026
Analysis · 6 min read · Aug 9, 2026The long view on DeepSeek: strategy, technology and what to watch02
2026-2028
Deploy proceeds of the ~$7B round into self-operated data centers built on Huawei Ascend chips to serve AI-agent workloads, deepening the domestic-silicon strategy proven by V4's Nvidia-free training run. Bloomberg reported on July 30, 2026 that DeepSeek is developing a roughly one-gigawatt facility in Ulanqab, Inner Mongolia, building part of it itself while leasing additional capacity from other providers, and targeting bringing at least part of the capacity online in late 2027 or early 2028. Bloomberg reported on September 4, 2026 that DeepSeek plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT chips at the facility for running (not training) its models, which would be among the largest known clusters of Chinese AI accelerators; Huawei's limited production capacity is expected to cap 950DT output in the low hundreds of thousands of units this year, so delivery could take more than a year, and DeepSeek has reportedly asked Beijing to help secure faster, larger allocations.
2026 and beyond
Follow through on the August 6, 2026 agreement disclosed alongside DeepSeek's 140.8M yuan (~$20.8M) strategic stake in Unitree Robotics' Shanghai IPO: a pact to jointly develop AI models for humanoid robots, with Unitree giving DeepSeek priority on model-training procurement and DeepSeek favoring Unitree for robots and embodied-AI applications. It is DeepSeek's first disclosed move into robotics, and the shares carry a 36-month lock-up. No shipped joint output has been disclosed yet.
Late 2026
TargetLate 2026 (filing); 2027 (listing)
Submit the domestic listing filing reported to be targeted for late 2026 toward a public debut on Shanghai's STAR Market in 2027. Reuters reported on September 9, 2026 that DeepSeek has tapped CITIC Securities to prepare the offering and aims to begin the listing process this year; South China Morning Post said CITIC is one of four underwriters. Deal size, timetable and target valuation are not set. The follow-on raise paused in late July after founder Liang Wenfeng's leaked investor remarks went viral resumed around August 6, 2026, and by late August DeepSeek was finalizing the round at roughly 50B yuan (~$7.4B) and a ~$74B pre-money valuation with returning backers Monolith Management, Shixiang Capital, Tencent and CATL. Sources had expected that raise to close by the end of August 2026; as of September 9, 2026 no public confirmation of a close has appeared, and no IPO filing has been announced.
V4.1-Pro undated
TargetV4.1-Pro undated; round targeting end of October 2026
No date announced
DeepSeek said V4.1-Flash will stand in for V4-Pro on the API from 04:00 UTC on September 14, 2026, with all deepseek-v4-pro requests routing to V4.1-Flash at Flash rates until V4.1-Pro launches. No launch date for V4.1-Pro has been announced.
03
DeepSeek said on September 30, 2026 via its official WeChat account that it has open-sourced Ascend-platform infrastructure matching its Nvidia-side stack, including TileLang plus DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Phoenix reprints said Huawei's team supported the work and that the companies jointly advanced a 128-card supernode based on Ascend 950. SCMP reported TileLang now lists Ascend 950 support, and DeepSeek's DeepGEMM-Ascend GitHub repository describes a matrix-multiplication library for Huawei Ascend NPUs.
tech.ifeng.com tech.ifeng.com www.scmp.com github.com + Huawei Investment & Holding Co., Ltd.
The Information reported on September 24, citing two people with direct knowledge, that DeepSeek's annualised revenue run rate has reached $1 billion, more than double a few months earlier. CEO Liang Wenfeng shared the figure with investors as the company targets a second funding round of 50 billion yuan ($7.45 billion) at a 500 billion yuan valuation by the end of October; Reuters said it could not immediately verify the report and could not reach DeepSeek. The same report said recent 2.3x to 4.5x model price increases helped the run-rate, and that Liang is still allocating more than 70% of compute to training.
DeepSeek posted an arXiv systems paper describing DeepSeek Elastic Compute, a production sandbox platform used for agent reinforcement-learning. The paper, with more than 130 authors including founder Liang Wenfeng as corresponding author, says a production unit of about 160 nodes serves about 3 million sandboxes a day, more than 380,000 concurrent sandboxes and more than 5,000 sandbox creations per second. Bloomberg reported the disclosure on September 25. The scale figures are DeepSeek's own and have not been independently measured; the paper is not an open-source release.
DeepSeek launched DeepSeek-V4.1-Flash on September 10 as a 552-billion-parameter mixture-of-experts model with a Causal Encoder-Decoder architecture, 8 billion active parameters on input and 16 billion on output, and native visual understanding. The API model name is deepseek-flash; V4-Flash and V4-Flash-Vision-Exp are retired, with compatibility aliases routing to V4.1-Flash. Starting at 04:00 UTC on September 14, all deepseek-v4-pro requests route to V4.1-Flash at Flash rates until V4.1-Pro launches, and new peak/off-peak pricing (off-peak at 50 percent of peak) took effect at 04:00 UTC on September 10.
CISA, NSA and the FBI warned on September 8, 2026 that China-based AI companies have run industrial-scale knowledge-distillation campaigns against US frontier models since at least late 2024. The agencies named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as extracting billions of tokens across millions of requests from Claude, GPT, Gemini and Grok variants, in violation of US providers' terms of service. They urged US labs to detect anomalous accounts, degrade suspected distillation traffic and share intelligence across providers.
Reuters reported on September 9, 2026 that DeepSeek has tapped CITIC Securities to prepare a listing on Shanghai's STAR Market and aims to start the IPO process this year, though deal size and timetable are not set. South China Morning Post said CITIC is one of four underwriters. The talks come as DeepSeek works through a follow-on round that Reuters has said would value the company at about 500 billion yuan, or roughly $75 billion.
DeepSeek plans to deploy at least 160,000 of Huawei's next-generation Ascend 950DT accelerators at the gigawatt-scale data center it is building in Ulanqab, Inner Mongolia, Bloomberg reported September 4, 2026, in what could become one of the largest known clusters of Chinese AI chips. Huawei's production capacity is expected to limit 950DT output to the low hundreds of thousands of units this year, so delivery could take more than a year, and DeepSeek has reportedly asked Beijing to help secure faster, larger allocations. The chips are intended to run DeepSeek's models rather than train them, with training workloads still reliant on whatever Nvidia hardware the company can obtain.
DeepSeek is finalizing a roughly 50 billion yuan ($7.4 billion) funding round that values the company at about $74 billion pre-money, up from just over $50 billion after its June 2026 round, with the raise expected to close by the end of August 2026. Returning backers include Monolith Management, Shixiang Capital, Tencent, and CATL, with proceeds earmarked for compute buildout and model research. The round is widely seen as preparation for a potential Shanghai STAR Market filing as early as late 2026, ahead of a targeted 2027 public debut.
Taiwanese threat-intelligence firm TeamT5 said Chinese state-affiliated hacking groups now run roughly twice as many attacks as before they began delegating tasks to DeepSeek and other open-weight AI models, Bloomberg reported August 24. Researchers obtained scripts and logs showing DeepSeek used across multiple attack stages, including exploit-code generation and target reconnaissance, and said the model's combination of strong performance, low cost and weaker safety guardrails than many Western models makes it attractive to threat actors.
Unitree Robotics began trading on Shanghai's STAR Market on August 19, becoming the first humanoid robot maker listed on a mainland China exchange. Shares opened 629% above the 150.80 yuan IPO price before paring gains to trade roughly 500-540% higher during the session, an opening market cap of about 62-66 billion dollars. The offering raised roughly 904 million dollars and was oversubscribed more than 8,000 times, with backing from strategic investors including DeepSeek.
DeepSeek released DeepSeek-V4-Flash-Vision-Exp on August 21, an experimental model that opens its API to image inputs alongside text for the first time, billed at the same per-token rate as the existing V4-Flash model. The model scored 27.3 on the Agents Last Exam benchmark, ahead of Anthropic's Opus 4.8 at 25.7, and DeepSeek says its vision pipeline uses roughly 90 KV-cache entries per image versus about 870 for Claude's vision system.
DeepSeek's new peak/off-peak billing for its V4-Pro and V4-Flash APIs took effect August 16, replacing flat per-token rates with time-of-day pricing. V4-Pro output rose from a flat $0.87 per million tokens to $3.96 per million at peak hours (01:00-04:00 and 06:00-10:00 UTC, half that off-peak), and V4-Flash output rose from $0.28 to $1.32 per million at peak, increases the company flagged in mid-August as it pursues a possible IPO.
DeepSeek moved V4-Pro out of preview to general availability across its app, web interface and API on August 12, focused on agentic tool use and multi-step workflows with a 1 million token context window. New peak and off-peak API pricing takes effect August 16, raising V4-Pro's peak output price to $3.96 per million tokens from a prior flat $0.87.
DeepSeek told users on August 6 that it will implement a significant price increase across its AI services, reversing the ultra-low pricing strategy that has powered its rapid growth, though it did not specify exact new figures. The Hangzhou lab's cheap V4-Flash model reportedly processed 8 trillion tokens in a single day on August 1, straining available compute and pushing the company to reconsider the pricing that touched off China's AI price war.
DeepSeek reopened its second external funding round around August 6, seeking close to $8 billion at a valuation near 500 billion yuan (about $74 billion), with Monolith Management, an early Moonshot AI backer, reportedly in talks to participate. The round had been paused in late July after leaked remarks from founder Liang Wenfeng frustrated some investors, and the new target valuation is up from the roughly $71 billion figure reported when talks first emerged in mid-July.
04
DeepSeek's flagship open-weight models: V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B / 13B active), both MIT-licensed with 1M-token context and hybrid thinking/non-thinking modes - reportedly trained on Huawei Ascend chips without Nvidia GPUs.
Free consumer assistant across iOS, Android and web, running the V4 family with Expert and Instant modes. Briefly the #1 free app on the US iOS App Store after R1's January 2025 launch and still one of the world's top AI chat platforms.
Developer platform exposing the V4 family through OpenAI- and Anthropic-compatible endpoints at aggressive prices (V4-Flash from $0.14/$0.28 per million input/output tokens), with cache-hit input discounts and, from the official V4 release, peak/off-peak pricing.
05
(mid-2026); doubling headcount across departments underway
The Information reported an annualised revenue run rate of $1 billion as of late September 2026, more than double from under $500 million a few months earlier. Reuters and PYMNTS could not verify independently; DeepSeek did not comment. This is a forward run rate, not $1 billion of booked trailing-twelve-month revenue. The Information separately said first-seven-month 2026 revenue was about 475 million yuan (about $70.7 million).
a follow-on targeting ~50B yuan (~$7.4B) at a ~$74B pre-money valuation was still being finalized as of September 8, 2026, past the end-of-August close sources had expected, with no public confirmation of a close
(QuestMobile, May 2026, published Jul 14 - the most recent published QuestMobile snapshot as of late August 2026); third-largest Chinese AI-native app behind ByteDance's Doubao (~382M MAU) and Alibaba's Qwen (~167M MAU)
Third among China's AI-native apps by MAU (QuestMobile, May 2026), behind Doubao and Qwen - the earlier ~89% 'dominant' estimate reflected the Jan 2025 R1-launch spike and no longer holds
V4.1-Flash peak/off-peak pricing took effect at 04:00 UTC on September 10, 2026; off-peak rates are 50 percent of peak. Starting 04:00 UTC on September 14, deepseek-v4-pro requests route to V4.1-Flash at Flash rates until V4.1-Pro launches. DeepSeek did not republish dollar amounts on the V4.1-Flash announcement. The prior table (in effect since August 16) listed V4-Flash and Vision-Exp at $0.22/$0.66 off-peak and $0.44/$1.32 peak per million cache-miss input/output tokens, and V4-Pro at $0.66/$1.98 off-peak and $1.32/$3.96 peak
(2.788M H800 GPU-hours, per DeepSeek's own Dec 2024 V3 technical report) vs ~$100M+ est. for GPT-4
(mid-2026); DeepSeek announced Jun 25, 2026 plans to double headcount across departments (pre-training data, server engineering, AI search, data-center ops, a new Harness agent team) following its $7.4B raise. No newer official headcount has been published as of September 8, 2026
The Information, citing two people with knowledge, said Liang told investors the annualised run rate reached $1 billion, more than double from under $500 million a few months earlier, helped by a 2.3-4.5x model price rise. Reuters could not verify and DeepSeek did not comment. This is a forward pace, not booked FY revenue. First-seven-month 2026 revenue was separately reported at about 475 million yuan. The follow-on round is still described as targeting 50 billion yuan at a 500 billion yuan valuation by the end of October, with no public close as of 26 Sep 2026.
Official table as of 26 Sep 2026: V4.1-Flash (deepseek-flash) cache-miss input/output $0.15/$0.60 off-peak and $0.30/$1.20 peak per million tokens; V4-Pro-0813 $0.66/$1.98 off-peak and $1.32/$3.96 peak. Off-peak is 50% of peak. Peak windows remain 01:00-04:00 and 06:00-10:00 UTC Monday-Friday excluding Chinese public holidays. Flash concurrency limit 2500; V4-Pro 500. Flash is cheaper than the prior V4-Flash $0.22/$0.66 off-peak table.
06
Operating status
Scaling; Ascend software stack open-sourced
On September 30, 2026 DeepSeek open-sourced TileLang and related compute and communication libraries for Huawei Ascend, with company language that Huawei supported a 128-card Ascend 950 supernode. DeepSeek-V4.1-Pro has still not launched on the API or as published weights. Official changelog latest dated entry remains 10 September: V4.1-Flash is live as deepseek-flash, and DeepSeek continues V4 Pro API service after 14 September. No public close of the reported 50 billion yuan follow-on has appeared.
Usage-based, per-token access to the V4 family via OpenAI- and Anthropic-compatible endpoints, now billed on official peak/off-peak rates (weekday peak windows 01:00-04:00 and 06:00-10:00 UTC) that took effect August 16, 2026.
The DeepSeek app and chat.deepseek.com are free - a deliberate user-growth strategy that leaves the API as the primary revenue source for now.
07
How past announced dates turned out.
1/2announced dates hit on time
Ship the official DeepSeek-V4 release
Target Aug 2026 · Resolved Aug 12, 2026
DeepSeek shipped V4-Pro-0813 to general availability across app, web and API on August 12, 2026, graduating it out of preview with a 1 million-token context window, up to 384,000-token output and a focus on agentic tool-use and multi-step workflows. New peak/off-peak API pricing took effect August 16, 2026, raising V4-Pro's peak output price to $3.96 per million tokens from the prior flat $0.87 rate.
Ship the official DeepSeek-V4 release
Target Jul 2026 · Resolved Aug 1, 2026
The mid-July 2026 target for a full V4 GA passed without one. On July 31, 2026 DeepSeek put only a retrained V4-Flash-0731 API into public beta (outperforming the V4-Pro preview on nine published agent/coding benchmarks), while the V4-Pro API and the app/web models were unchanged. Chinese tech media reported the complete rollout, including peak/off-peak pricing, sliding to a possible August 10-20, 2026 window tied to closed testing of DeepSeek's new 'Harness' agentic coding tool.
08
2026
Open-sources Ascend TileLang stack with Huawei support
On September 30, 2026 DeepSeek said on WeChat that it open-sourced Ascend-platform infrastructure matching its Nvidia-side tools: TileLang plus DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. Phoenix reprints said Huawei's team supported the work and that the sides jointly advanced a 128-card supernode based on Ascend 950. DeepSeek's DeepGEMM-Ascend GitHub repository describes a matrix-multiplication library for Huawei Ascend NPUs.
Keeps V4-Pro on the API after reversing the Sept 14 retirement
Official API docs dated 10 Sep 2026 say DeepSeek will continue providing API services for DeepSeek V4 Pro after 14 Sep 2026, with billing unchanged, reversing the V4.1-Flash launch post that had said all deepseek-v4-pro requests would route to Flash at Flash rates from 04:00 UTC on 14 Sep until V4.1-Pro launches. The live pricing table still lists deepseek-v4-pro (V4-Pro-0813) beside deepseek-flash. V4-Flash and V4-Flash-Vision-Exp remain retired aliases. No V4.1-Pro date has been announced.
Publishes DSec agent-training sandbox paper
DeepSeek posted an arXiv systems paper on DeepSeek Elastic Compute, a production sandbox platform with FnCall, container, microVM and full-VM backends. The paper says a production unit of about 160 nodes serves about 3 million sandboxes a day, more than 380,000 concurrent sandboxes and more than 5,000 sandbox creations per second. Bloomberg reported the disclosure on September 25. The figures are DeepSeek's own; the paper is not an open-source release.
The Information reports $1 billion annualised revenue run rate
The Information reported, citing two people with knowledge, that DeepSeek's annualised revenue run rate reached $1 billion. Reuters could not independently verify the figure and could not reach the company. Liang told investors the company is targeting a 50 billion yuan second round at a 500 billion yuan valuation by the end of October.
Ships DeepSeek-V4.1-Flash and starts retiring V4-Pro
On September 10, 2026 DeepSeek launched DeepSeek-V4.1-Flash, a 552-billion-parameter mixture-of-experts model with a Causal Encoder-Decoder architecture (8 billion active parameters on input, 16 billion on output) and native visual understanding. The API model is deepseek-flash; V4-Flash and V4-Flash-Vision-Exp are retired, with compatibility aliases routing to V4.1-Flash. New peak/off-peak pricing took effect at 04:00 UTC on September 10, with off-peak rates at 50 percent of peak. Starting at 04:00 UTC on September 14, all deepseek-v4-pro requests route to V4.1-Flash at Flash rates until V4.1-Pro launches.
Taps CITIC Securities for a Shanghai STAR Market IPO
Reuters reported on September 9, 2026 that DeepSeek has tapped CITIC Securities to prepare an initial public offering on Shanghai's STAR Market and aims to begin the listing process this year, though deal size, timetable and target valuation are not set. South China Morning Post said CITIC is one of four underwriters. The engagement is a typical pre-filing tutoring step for a mainland listing and comes as DeepSeek works through a follow-on round that Reuters has said would value the company at about 500 billion yuan (roughly $75 billion).
Named in CISA-NSA-FBI industrial-scale distillation advisory
On September 8, 2026 CISA, NSA and the FBI warned that China-based AI companies have run industrial-scale knowledge-distillation campaigns against US frontier models since at least late 2024. The agencies named DeepSeek, along with Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, as extracting billions of tokens across millions of requests from Claude, GPT, Gemini and Grok variants in violation of US providers' terms of service. The advisory said the campaigns form the core of China's AI development strategy and urged US labs to detect anomalous accounts, degrade suspected distillation traffic and share intelligence.
Open-sources DeepSeek-V4-Flash-Vision-Exp weights
On August 31, 2026 DeepSeek published MIT-licensed weights for DeepSeek-V4-Flash-Vision-Exp on Hugging Face, ten days after opening the same experimental multimodal model on its API. The ~305B-parameter checkpoint (about 168GB across 48 Safetensors files) lets developers run the vision model locally rather than only through DeepSeek's billed API, which charges Vision-Exp at the same peak/off-peak rates as text-only V4-Flash.
2026
Open-sources DeepSeek Harness, an agentic coding runtime rivaling Claude Code
On August 13, 2026, alongside V4-Pro's general-availability launch, DeepSeek releases DeepSeek Harness v0.1 as open-source software under an MIT license - a plugin-first agent runtime for assembling coding agents from replaceable models, tools, skills, sessions, and sandboxes, positioned as a direct alternative to Anthropic's Claude Code. The developer preview passes roughly 95,000 GitHub stars within about two days of release.
Ships DeepSeek-V4-Flash-Vision-Exp, its first multimodal model
On August 21, 2026 DeepSeek releases DeepSeek-V4-Flash-Vision-Exp, an experimental model that opens its API to image inputs alongside text for the first time, billed at the same per-token rate as V4-Flash. On the Agents Last Exam benchmark it scores 27.3 versus 25.7 for Anthropic's Opus 4.8, and DeepSeek says its vision pipeline uses roughly 90 KV-cache entries per image versus about 870 for Claude's vision system.
V4-Pro reaches general availability
On August 12, 2026 DeepSeek ships V4-Pro-0813 to general availability across its app, web interface and API, exiting preview with a 1 million-token context window, up to 384,000-token output and a focus on agentic tool-use and multi-step workflows. New peak/off-peak API pricing takes effect August 16, 2026, raising V4-Pro's peak output price to $3.96 per million tokens from the prior flat $0.87 rate.
Follow-on funding round resumes at ~$74B valuation
DeepSeek reopens its paused follow-on funding round around August 6, 2026, seeking close to $8B at a valuation near 500B yuan (~$74B), with Monolith Management - an early Moonshot AI backer - reportedly in talks to participate. The round had been paused since July 25 after leaked remarks from founder Liang Wenfeng frustrated some investors.
Takes strategic stake in Unitree, signs embodied-AI pact
DeepSeek's parent entity becomes a strategic placement investor in Unitree Robotics' Shanghai STAR Market IPO, buying 933,400 shares (~141M yuan, ~$20.8M, under a three-year lock-up) alongside Tencent and several state-backed institutions. The two companies agree to co-develop AI models and embodied-intelligence technology and to give each other priority access to robot procurement and AI model services - DeepSeek's first disclosed move into humanoid robotics.
V4-Flash-0731 update posts major agentic gains
On July 31, 2026 DeepSeek publishes DeepSeek-V4-Flash-0731, a re-post-trained update to its 284B-total/13B-active-parameter model that posts large agentic and coding benchmark gains over its own V4-Pro despite far fewer active parameters, and moves the V4-Flash API into public beta with support for the Responses format used by coding tools like Codex.
IPO preparation and talks for a further ~$1.5B
In mid-July 2026 Bloomberg and the FT report DeepSeek is in talks to raise a further ~$1.5B at a ~$71B valuation and is preparing a mainland China IPO filing this year for a potential 2027 listing.
Record ~$7B maiden funding round at ~$52B valuation
In June 2026 DeepSeek closes its first-ever external round - roughly ¥51B (~$7B), the largest single AI funding round in China's history - at a ~$52B post-money valuation, with Tencent, CATL, JD.com, NetEase and a state AI fund participating, mostly via non-voting stakes that keep Liang Wenfeng in control.
DeepSeek-V4 family debuts, trained on Huawei silicon
On April 24, 2026 DeepSeek releases previews of V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B / 13B active) - MIT-licensed open-weight models with 1M-token context, reportedly trained on Huawei Ascend chips with no Nvidia GPUs, a landmark for China's domestic-compute push.
Anthropic alleges industrial-scale distillation
In February 2026 Anthropic publicly accuses DeepSeek - along with Moonshot AI and MiniMax - of harvesting Claude outputs for model distillation through roughly 24,000 fraudulent accounts and 16M+ exchanges, sharpening the US–China AI dispute; DeepSeek is also barred from US defense and intelligence systems under the FY2026 NDAA.
2025
V3.1 and V3.2 fold reasoning into hybrid models
DeepSeek ships V3.1 in August 2025 (hybrid thinking/non-thinking modes) and V3.2 with its Sparse Attention architecture on December 1, 2025 - absorbing the R-series reasoning line into the main model family after the long-rumored R2 never ships.
R1 release triggers a $600B Nvidia rout
On January 20, 2025 DeepSeek releases the MIT-licensed R1 reasoning model. Within a week its free app tops the US iOS App Store and, on January 27, Nvidia falls ~18% - erasing roughly $600B, the largest single-company one-day market-value loss in US history - in what is dubbed AI's 'Sputnik moment.'
2024
DeepSeek-V3 trained for a claimed ~$6M
In December 2024 DeepSeek releases V3, a frontier-class mixture-of-experts model it says was trained for roughly $6M - versus an estimated ~$100M for GPT-4 - establishing its extreme cost-efficiency playbook.
DeepSeek-V2 ignites China's LLM price war
DeepSeek releases V2 in May 2024 at disruptively low API prices, earning the nickname 'the Pinduoduo of AI' and forcing Chinese rivals including Alibaba, Baidu and Tencent to slash their own model pricing.
2023
09
$14.4Bdisclosed across 2 rounds with a stated amount
Internal funding (High-Flyer)
2023–2026
Undisclosed (wholly funded by High-Flyer)
High-Flyer / Liang Wenfeng (84% founder stake as of May 2024)
Maiden external round
Jun 2026
Tencent, CATL, JD.com, NetEase, Monolith, Loyal Valley Capital, China's National AI Industry Investment Fund; ¥20B from Liang Wenfeng
Follow-on (resumed)
Jul-Aug 2026
~50B yuan (~$7.4B) at a ~$74B pre-money valuation; DeepSeek paused signing of new agreements on Jul 25, 2026 after founder Liang Wenfeng's leaked investor remarks went viral, then reopened the round around Aug 6, 2026 and by late August was finalizing the raise with returning backers Monolith Management, Shixiang Capital, Tencent and CATL. Sources had expected a close by the end of August 2026; as of September 8, 2026 no public confirmation of a close has appeared
Monolith Management, Shixiang Capital, Tencent, CATL
