Skip to content

DeepSeek

Dossier checked 28 Sep 2026

What it sells now

Newest frontier release

DeepSeek-V4.1-Flash-1017

17 Oct 2026 · the lab’s announcement

Re-post-trained V4.1-Flash checkpoint, live on the web, app and API and billed at the Flash price, with open weights

Flagship list price, per million tokens

$1.32 in $3.96 out

DeepSeek-V4-Pro, cache-miss input, peak hours, released 13 Aug 2026 · pricing page

V4.1-Flash (10 Sep) is newer and cheaper, at $0.30 in and $1.20 out at peak, and beats V4-Pro on most of DeepSeek’s own agent benchmarks; V4-Pro is priced here as the largest model. Its 17 Oct retrain, V4.1-Flash-1017, is billed at the same Flash price.

Weights

Open flagship

Full, under MIT: V4-Pro, V4-Flash, V4.1-Flash, R1, and V3 from the March 2025 update on. The original V3 and earlier models used a custom DeepSeek licence. Source

Where you can buy the flagship

Its other models, not the flagship

Source

6 releases in the year to 30 Oct 2026, 9 the year before · 21 in all since 2023

Our read, not a fact

In the model

The question we ask of every lab: is most of its value in the model, which gets cheaper fastest, or in something other work is built on? Our answer: in the model itself.

DeepSeek lives at the sharp end of the output layer: open weights, frontier-adjacent quality, at a fraction of the closed labs’ prices. The durable asset is the efficiency itself, the ability to keep matching the frontier cheaply enough to reset everyone’s price expectations.

What would change our readThe cost advantage closes as rivals copy the efficiency tricks, or open weights stop tracking the closed frontier.

01 · What it runs on, and the money

Trains and serves on press report

Disclosed only for an earlier generation: DeepSeek-V3 was trained on NVIDIA H800 GPUs, per its own technical report, and the V4.1 report names no training hardware. A Huawei-led team said in June 2026 it had post-trained V4-Pro on about 1,000 Ascend 910C chips. Bloomberg reported in September 2026 that DeepSeek plans at least 160,000 Ascend 950DT chips for a gigawatt-scale inference site in Inner Mongolia. Source

What it depends on press reportFrontier training silicon needs clearance from two governments at once, a US export licence for the NVIDIA part and a Chinese import approval for the same part, while Beijing directs routine inference onto domestic accelerators. Source

Biggest risk press reportIts frontier training hardware needs US export authorisation and Chinese state approval at once, and the domestic alternative has shown only post-training and inference at V4 scale, not pre-training. Source

Raised press report
About RMB 50B (~$7.4B) in its first outside round, closed by July 2026: about RMB 20B from Liang Wenfeng himself, the rest from Tencent, CATL, NetEase, JD.com and the National AI Industry Investment Fund. A second round of similar size is reported to be closing. Source
Valuation press report
About US$51.8B post-money (Jul 2026); the second round is reported at a valuation of about US$74B Source
Revenue press report
Not disclosed. About $1B annualised, a figure the founder reportedly gave investors in September 2026 Source
Structure press report
Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Private and unlisted, grown out of the hedge fund High-Flyer, with founder Liang Wenfeng holding the large majority. Source

NotesSince 16 Aug 2026 the API bills peak and off-peak hours separately, with off-peak at half the peak rate (peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays); the price shown is the peak rate. What V4 and V4.1 were pre-trained on is unknown. The widely quoted 2.788M H800 GPU-hours and $5.5M training cost belong to V3, not V4.

02 · Releases

data through 30 Oct 2026

  1. DeepSeek 19 DeepSeek-V4.1-Flash-101717 Oct 2026

Each dot is a frontier release at its date. Click one to find it in the full log. Every lab on one chart

The landmarks

  1. Apr 2026 DeepSeek-V4 (Pro/Flash) Million-token context preview Open1M ctxFlagship
  2. Jan 2025 DeepSeek-R1 Open reasoning model that set off a global market shock OpenReasoningMarket shock
  3. Dec 2024 DeepSeek-V3 671B MoE trained at remarkably low cost Open671B MoE
  4. May 2024 DeepSeek-V2 Cheap MoE that triggered China’s price war OpenMoEPrice war
The full release logall 21, newest first
  1. 2026
  2. 17 Oct 2026 DeepSeek-V4.1-Flash-1017 Frontier Re-post-trained V4.1-Flash checkpoint, live on the web, app and API and billed at the Flash price, with open weights Source
  3. 10 Sep 2026 DeepSeek-V4.1-Flash Frontier First model of a new architecture family, on the API Source
  4. 13 Aug 2026 DeepSeek-V4-Pro (GA) Frontier General release of the V4-Pro that first shipped in April Source
  5. 31 Jul 2026 DeepSeek-V4-Flash Frontier Public beta of the Flash tier, re-trained from the April preview on the same architecture Source
  6. Apr 2026 DeepSeek-V4 (Pro/Flash) Frontier Million-token context preview
  7. 2025
  8. 1 Dec 2025 DeepSeek-V3.2 Frontier Plus a Speciale variant Source
  9. 29 Sep 2025 DeepSeek-V3.2-Exp Frontier Sparse-attention experiment, the default model on the app, web and API from launch Source
  10. 22 Sep 2025 DeepSeek-V3.1-Terminus Frontier Final V3.1 update, on the app, web and API and as open weights Source
  11. Aug 2025 DeepSeek-V3.1 Frontier Hybrid reasoning update
  12. May 2025 DeepSeek-R1-0528 Frontier Major R1 upgrade
  13. 24 Mar 2025 DeepSeek-V3-0324 Frontier Named V3 checkpoint behind deepseek-chat, open weights under MIT Source
  14. Jan 2025 DeepSeek-R1 Frontier Open reasoning model that set off a global market shock
  15. 2024
  16. Dec 2024 DeepSeek-V3 Frontier 671B MoE trained at remarkably low cost
  17. 10 Dec 2024 DeepSeek-V2.5-1210 Frontier Named final V2.5 checkpoint, on the web and API and as open weights Source
  18. 20 Nov 2024 DeepSeek-R1-Lite-Preview Frontier First DeepSeek reasoning model, free on chat.deepseek.com; no API or weights yet Source
  19. 5 Sep 2024 DeepSeek-V2.5 Frontier Merged the chat and coder models; on the web and API and as open weights Source
  20. 28 Jun 2024 DeepSeek-V2-0628 Frontier Named V2 checkpoint behind deepseek-chat from 28 Jun, released as open weights (DeepSeek-V2-Chat-0628) on 18 Jul Source
  21. Jun 2024 DeepSeek-Coder-V2 Strong open coding model
  22. May 2024 DeepSeek-V2 Frontier Cheap MoE that triggered China’s price war
  23. 2023
  24. Nov 2023 DeepSeek LLM Frontier 67B base and chat models
  25. Nov 2023 DeepSeek Coder First model, code-specialised

03 · When to reach for them

practical read

Cheap open agent model
Reach for V4.1-Flash when you want an open, multimodal agent model at $0.30 in and $1.20 out per million tokens (Sep 2026).
Largest open model to host
Pick V4-Pro for a 1.6T-parameter mixture-of-experts model you can download under MIT and host yourself (Aug 2026).
Cost-sensitive scale
Use DeepSeek where token economics decide the build, and run batch jobs off-peak at half the peak rate (since Aug 2026).

04 · Weights, across all ten

the flagship

Open flagship
DeepSeek Mistral AI Alibaba (Qwen) Moonshot AI Zhipu (Z.ai)
Closed flagship, some open
OpenAI Google Meta xAI
Closed
Anthropic

05 · Essays that read DeepSeek

from the archive

The Model Is Not the Moat If frontier capability keeps centralising, the durable edge shifts outward into trust, workflow fit, and the surrounding package. You're Not Comparing Models. You're Comparing Contracts. Agent benchmarks don't measure models. They measure contracts. Two teams running the same model can publish different scores, and both can be honest. The Displacement Rate Audit A five-minute scoring tool for any product, position, architecture, business model, or career bet. Find out what still works after the environment changes.

06 · About DeepSeek

at a glance

Founded
2023
HQ
Hangzhou, China
Leadership
Liang Wenfeng
Parent
High-Flyer
Also builds
DeepSeek app, DeepSeek Harness (preview)
Known for
Low-cost open models

Founded July 2023 by Liang Wenfeng, spun out of the Chinese quant fund High-Flyer. Headquartered in Hangzhou, it is known for cost-efficient open-weight reasoning models; the R1 release in January 2025 triggered a global reaction over China’s frontier capabilities.