Historical model page: This page covers DeepSeek V3 as a historical release from December 2024. It is kept for reference, architecture notes, release context, and timeline research. It is not a live model-selection guide, pricing guide, or API implementation guide.
Last reviewed: April 24, 2026
DeepSeek V3 was introduced in December 2024 as a major milestone in the DeepSeek model lineage. It brought the V3 generation into the hosted API at that point in time, established a 671B-parameter Mixture-of-Experts architecture with 37B activated parameters per token, and combined long-context support with a broad open-weight release.
This page is intentionally written as a historical overview. It is useful for readers researching the DeepSeek timeline, the original V3 open-weight release, or the point at which the hosted API moved beyond the V2.5 line. It should not be treated as a reference for the live DeepSeek API surface.
Where DeepSeek V3 Fits in the Timeline
DeepSeek V3 matters historically because it marked the transition from the V2.5 era to the V3 line. Later releases continued the V3 family before DeepSeek moved into newer generations. The timeline below is included to explain model history, not to recommend any older model name for new API work.
| Date | Milestone | Why it matters historically |
|---|---|---|
| December 26, 2024 | DeepSeek V3 release | deepseek-chat was upgraded to DeepSeek V3 at that time, bringing the V3 line into the hosted API. |
| March 25, 2025 | DeepSeek-V3-0324 | The V3 line received a follow-up update with stronger reasoning, stronger front-end development skills, and smarter tool-use capabilities while preserving the broader V3 generation context. |
| August 21, 2025 | DeepSeek V3.1 | DeepSeek V3.1 continued the evolution of the V3 family after the original V3 release. |
| December 1, 2025 | DeepSeek V3.2 | DeepSeek V3.2 became a later V3-generation release after V3.1. It should be understood as a later point in the V3 lineage, not as part of the original December 2024 V3 release. |
| April 24, 2026 | DeepSeek V4 Preview | DeepSeek V4 Preview marked a newer generation after the V3 line. This is why this DeepSeek V3 page should be read as an archive, not as a live API reference. |
What DeepSeek V3 Introduced
Compared with the earlier V2.5 line, DeepSeek V3 introduced a new large-scale MoE model family with several historically important characteristics:
- 671B total parameters with 37B activated per token: DeepSeek V3 used a large Mixture-of-Experts design that activated only part of the model on each token.
- 128K context window: the V3 model card listed a 128K context window for both the Base and Chat releases.
- 14.8T-token pretraining: the model card described pretraining on 14.8 trillion diverse, high-quality tokens.
- MLA and DeepSeekMoE: the release emphasized Multi-head Latent Attention and the DeepSeekMoE architecture as core parts of the system design.
- Auxiliary-loss-free load balancing and multi-token prediction: DeepSeek positioned these as key technical changes in the V3 generation.
- Base and Chat releases: the model card listed both DeepSeek-V3-Base and DeepSeek-V3.
Model Snapshot
| Release window | December 2024 |
|---|---|
| Model family | DeepSeek V3 line |
| Total parameters | 671B |
| Activated parameters | 37B |
| Context length | 128K |
| Pretraining scale | 14.8T tokens |
| Core architectural terms | Mixture-of-Experts, Multi-head Latent Attention, DeepSeekMoE |
| Historical API mapping at release | deepseek-chat was upgraded to DeepSeek V3 in December 2024. |
| Archive status | This page documents the original DeepSeek V3 release and its role in the model timeline. It is not a live API model recommendation. |
| Model card licensing note | The code repository was listed under MIT, while the model card stated that DeepSeek-V3 Base and Chat were subject to the model license and supported commercial use. |
Open-Weight Release vs Hosted API
It is important not to blur two different contexts. The original DeepSeek V3 release included open-weight model distribution, technical documentation, and self-hosted deployment paths. The hosted API, by contrast, exposed DeepSeek V3 through the historical deepseek-chat alias starting in December 2024.
That API mapping describes the release period of DeepSeek V3. It should not be copied into new API examples as a model recommendation. For implementation, pricing, or model-selection decisions, use the dedicated live API, pricing, and model pages rather than this archive.
Why DeepSeek V3 Still Matters Historically
DeepSeek V3 remains an important reference point because it established the first V3-generation model line, brought the V3 architecture into the hosted API during its release period, and set the base that later advanced through V3-0324, V3.1, and V3.2. Readers comparing DeepSeek model generations often need this page to understand where that transition began.
It is also historically relevant for readers looking at the original V3 open-weight release, the first appearance of this 671B/37B configuration, or the technical shift toward MLA, DeepSeekMoE, and multi-token prediction in the V3 line.
Who Should Still Read This Page
- Readers comparing DeepSeek model releases over time
- Builders tracing how
deepseek-chatevolved across the V3 line - Researchers looking for the original DeepSeek V3 model release context
- Users arriving from older bookmarks, release notes, or benchmark discussions
- Writers checking the historical relationship between V2.5, V3, V3.1, V3.2, and later DeepSeek releases
How to Use This Historical Page
Use this page when you need background on the original DeepSeek V3 release, its architecture, its historical API mapping, or its role in the broader DeepSeek timeline.
Do not use this archive as the only source for implementation decisions. For live API usage, model IDs, pricing, context limits, and availability, check the Models hub, the API guide, the pricing page, and newer release pages such as DeepSeek V4.
If the goal is self-hosting or technical research on the original V3 release, use the original DeepSeek V3 model card and release materials for the exact checkpoint rather than assuming that the December 2024 V3 release matches later hosted API behavior.
Final Takeaway
DeepSeek V3 remains historically important because it marked the start of the V3 generation, introduced a large-scale 671B/37B MoE model with 128K context, and moved the hosted API beyond the V2.5 line during its release period.
This page should be read as a historical DeepSeek V3 archive. It explains what DeepSeek V3 was, why it mattered, and how it fits into the model timeline, without presenting it as the latest, default, or recommended hosted DeepSeek model.