CURATED TIMELINE · EDITORIAL EDITION

DeepSeek Before the Breakout

The January 2025 sensation did not appear overnight. Trace the open-weight code models, sparse experts, GRPO, MTP, V2 efficiency work, R1 preview, V3, the free app, and the App Store shock — while separating reported training-run cost from total R&D and open weights from full open source.

Timeline overview

At a glance

Hover for the event briefScroll horizontally

2023.112025.01

11.02DeepSeek Coder 01.11DeepSeekMoE 02.05DeepSeekMath / GRPO 05.07DeepSeek-V2 11.20R1-Lite-Preview 12.26DeepSeek-V3 01.27DeepSeek goes mainstream

10 nodes2023.11—2025.01 · Editorially curated · Updated 2026-07-15

Editorial thread

The timeline

Reviewed event briefs and original editorial context, ordered to show how the story changed over time.

01
2023-11-02Open-source release
DeepSeek starts with an open-weight code model
DeepSeek releases code models from 1.3B to 33B parameters, making open weights and code part of its product strategy from the outset.
Editorial context
The origin story is developer credibility, not an overnight consumer hit. High-Flyer supplied capital and early compute, but DeepSeek operated as a dedicated AI research effort. Calling it a quant fund's casual side project understates the sustained work that followed.

Its first model set the distribution strategy: release weights, code, and technical material to developers before building a mass-market chat product.
Read event
02
2024-01-11Paper & weights release
DeepSeek publishes the sparse architecture behind its later models
DeepSeekMoE uses fine-grained experts and shared expert isolation to improve sparse-model efficiency, with 16B weights and training code released.
Editorial context
Efficiency became a research program. DeepSeekMoE split experts more finely and isolated shared knowledge so only a small slice of a large model had to run for each token. That line later fed directly into V2 and V3.

Export controls are part of the operating context, but public evidence does not prove that each architectural idea was caused by them. The defensible claim is that efficiency became central under constrained hardware conditions.
Read event
03
2024-02-05Paper & weights release
DeepSeekMath introduces the GRPO method later used by R1
DeepSeek releases math models and Group Relative Policy Optimization, creating a direct algorithmic predecessor to R1's reinforcement-learning recipe.
Editorial context
R1's algorithmic prehistory is public. DeepSeekMath introduced GRPO in February 2024 and shipped Base, Instruct, and RL checkpoints. R1 later used the same family of reinforcement-learning methods.

By replacing PPO's separate critic with a group-relative baseline, GRPO reduced memory overhead and fit domains such as math and code where answers can be automatically checked.
Read event
04
2024-04-30Industry event
Meta researchers publish multi-token prediction
Meta FAIR researchers propose training language models to predict several future tokens at once, work that later informs DeepSeek-V3's MTP design.
Editorial context
Open research moved into production in about eight months. Meta FAIR's multi-token prediction paper appeared in April 2024; DeepSeek-V3 cited it in December.

V3 did not copy the paper unchanged. It converted parallel prediction heads into a sequential module that preserved causal dependencies. The more useful lesson is how quickly a public idea could be adapted and engineered into a frontier system.
Read event
05
2024-05-07Open-source release
DeepSeek releases the efficiency-focused DeepSeek-V2
DeepSeek releases a 236B mixture-of-experts model that combines sparse activation with latent attention to cut training and inference costs.
Editorial context
V2 is where the architecture converged. DeepSeek combined its sparse expert design with Multi-head Latent Attention, shrinking the inference cache while improving throughput. Developer-reported comparisons showed large savings versus DeepSeek 67B.

V3 was therefore an extension of a tested stack, not a sudden breakthrough assembled in December.
Read event
06
2024-11-20Model preview
DeepSeek previews its reasoning line before V3 ships
R1-Lite-Preview exposes long-form reasoning on the web and promises an open model and API, showing that R1 was already underway before V3's release.
Editorial context
R1 was visible before V3 shipped. R1-Lite-Preview exposed long-form reasoning more than a month before the V3 launch and promised an open model and API.

V3's own report also said reasoning data came from an internal R1 series. The clean chronology is parallel development followed by adjacent public releases, not a reasoning model built from scratch in 25 days.
Read event
07
2024-12-26Open-source release
DeepSeek-V3 brings an efficiency-first open model to the frontier race
DeepSeek releases a 671B MoE model with 37B active parameters, open weights, low API prices, and unusually detailed efficiency disclosures.
Editorial context
V3 supplied the engineering headline. One release combined 671B total parameters, 37B active parameters, open weights, low API pricing, and strong developer-reported benchmarks.

The famous $5.576 million figure is a conversion of 2.788 million H800 hours for the reported final training run. It is not DeepSeek's total R&D cost and excludes earlier experiments, staff, data work, and cluster capital. The release is best described as open-weight: the repository code is MIT, while model weights use the DeepSeek Model License.
Read event
08
2025-01-15Product launch
DeepSeek launches a free consumer app before R1
DeepSeek puts V3, web search, a reasoning mode, and file analysis into a free mobile app five days before the R1 release.
Editorial context
Distribution arrived before the flagship reasoning release. Five days before R1, DeepSeek announced a free app with V3, search, Deep-Think, file upload, and synchronized history.

That meant R1 entered an existing consumer funnel. Millions of people could try the model without touching a repository, local runtime, or API console.
Read event