DeepSeek Release Timeline¶
Verified through: 2026-07-23. This timeline covers named first-party DeepSeek model families, open checkpoints, and material API model updates. It excludes third-party quantizations, community conversions, serving-framework ports, kernels, infrastructure projects, and ordinary bug fixes.
Read this page together with the model training token ledger. A release date does not disclose a training recipe, and a product alias does not necessarily identify a new checkpoint.
Dating and scope rules¶
DeepSeek's early repositories do not all contain a formal release log. This page therefore labels the event type:
- release: an official first-party announcement or repository news entry;
- weights: the official registry exposes the checkpoint;
- report: the first technical report became public;
- API update: a hosted alias changed its backing model; and
- temporary variant: a time-limited endpoint or mode, not a new base family.
For early weights without an announcement date, the date comes from the
official Hugging Face registry's createdAt metadata. That is artifact evidence
of availability, not a claim about a press launch. Where a report and weights
appeared on different dates, both are shown.
Main text and reasoning lineage¶
| Date | Model or event | Event type | Training-token status | What changed |
|---|---|---|---|---|
| 2023-11-29 | DeepSeek-LLM 7B / 67B | weights/repository | 2T [D] | dense bilingual base/chat family; registry, report |
| 2024-01-08–09 | DeepSeekMoE 16B Base / Chat | weights | 2T [D] | first public fine-grained/shared-expert MoE family; base and chat registry records appeared on consecutive days; base registry, repository/report |
| 2024-05-06 | DeepSeek-V2 | release | 8.1T [D] | 236B-total / 21B-active MoE, Multi-head Latent Attention (MLA), 128K; official repository news |
| 2024-05-16 | DeepSeek-V2-Lite | release | 5.7T [D] | 16B-total / 2.4B-active smaller V2 family; official repository news |
| 2024-05-17 | DeepSeek-V2-0517 | API update | inherits V2 | first dated V2 hosted update in the current change log; change log |
| 2024-06-28 | DeepSeek-V2-0628 | API update | added amount unknown [U] | improved reasoning and role-play behavior; change log |
| 2024-09-05 | DeepSeek-V2.5 | release/API update | added amount unknown [U] | combines V2-0628 conversational and Coder-V2-0724 coding capabilities; announcement |
| 2024-11-20 | DeepSeek-R1-Lite-Preview | web-only preview | training exposure unknown [U] | first public reasoning preview with visible reasoning and inference scaling; no weights or API were released; official announcement |
| 2024-12-10 | DeepSeek-V2.5-1210 | API update | added amount unknown [U] | math, coding, writing, file and webpage handling update; announcement |
| 2024-12-26 | DeepSeek-V3 | release | 14.8T [D] | 671B-total / 37B-active MoE, FP8 training, Multi-Token Prediction (MTP), loss-free routing; repository, report |
| 2025-01-20 | DeepSeek-R1-Zero and DeepSeek-R1 | release | inherit V3-Base; new broad total unknown [U] | no-SFT RL experiment plus production cold-start/SFT/RL pipeline; official release |
| 2025-01-20 | six DeepSeek-R1-Distill checkpoints | same-day release | post-training only | Qwen2.5- and Llama-based 1.5B, 7B, 8B, 14B, 32B, and 70B students; repository |
| 2025-03-24 | DeepSeek-V3-0324 | release/API update | added amount unknown [U] | post-training refresh for reasoning, code, writing, and function calling; change log |
| 2025-05-28 | DeepSeek-R1-0528 | release/API update | added amount unknown [U] | additional post-training compute and stronger reasoning; change log |
| 2025-08-21 | DeepSeek-V3.1 | release/API update | V3.1-Base adds about 839B CPT [D] | hybrid thinking/non-thinking, tool-agent training, 128K; change log, base model card |
| 2025-09-22 | DeepSeek-V3.1-Terminus | release/API update | added amount unknown [U] | language consistency and agent behavior update; change log |
| 2025-09-29 | DeepSeek-V3.2-Exp | release/API update | about 945.8B sparse conversion exposure [D] | experimental DeepSeek Sparse Attention (DSA); repository/report |
| 2025-12-01 | DeepSeek-V3.2 | release | added broad amount unknown [U] | production sparse-attention and agent/reasoning model; official transparency page, report |
| 2025-12-01 | DeepSeek-V3.2-Speciale | temporary variant | not a separate base count | high-compute reasoning endpoint released alongside V3.2 for a limited period; change log |
| 2026-04-24 | DeepSeek-V4-Flash / V4-Pro | preview release | 32T / 33T [D] | million-token context; new attention, hyper-connection, optimizer, post-training, and agent systems; official release, transparency page, immutable report snapshot |
Mainline interpretation¶
DeepSeek-LLM
-> DeepSeekMoE -> V2 / V2-Lite
-> V2.5
-> R1-Lite-Preview
-> V3
-> R1-Zero / R1 / R1-Distill
-> V3-0324
-> V3.1 -> V3.1-Terminus -> V3.2-Exp -> V3.2
-> V4-Flash / V4-Pro
This is a capability and checkpoint lineage, not a statement that every arrow is ordinary continued pretraining. V2.5 is a capability combination with an undisclosed merge/training recipe; R1 is post-training over V3-Base; V4 is a new reported pretraining program.
Code, mathematics, and theorem proving¶
| Date | Model or event | Event type | Token/accounting status | Evidence |
|---|---|---|---|---|
| 2023-10-28 | DeepSeek-Coder v1 weights | weights | 2T fresh code-model training [D] | registry, repository |
| 2024-01-25 | DeepSeek-Coder paper and Coder-v1.5 weights | report/weights | v1 is 2T; exact v1.5 lineage total is not cleanly restated | paper, v1.5 registry |
| 2024-02-05 | DeepSeekMath 7B | release/weights/report | +500B CPT over Coder-v1.5 [D] | repository, report |
| 2024-05-23 | DeepSeek-Prover V1 report | report | base inherited; added broad total unknown [U] | report |
| 2024-06-14 | DeepSeek-Coder-V2-0614 | API update/weights | V2 intermediate 4.2T + 6T CPT = 10.2T branch [D] | change log, registry |
| 2024-06-17 | DeepSeek-Coder-V2 report | report | same 10.2T branch | report |
| 2024-07-24 | DeepSeek-Coder-V2-0724 | API update | added amount unknown [U] | change log |
| 2024-08-15 | DeepSeek-Prover-V1.5 | weights/repository/report | 9B formal-proof SFT tokens; inherited base | repository, report |
| 2024-08-16 | DeepSeek-Prover V1 weights | weights | inherited base; added broad total unknown [U] | registry |
| 2025-04-30 | DeepSeek-Prover-V2 7B / 671B | weights/repository/report | based on Prover-V1.5 / V3-Base; added broad total unknown [U] | repository, report |
| 2025-11-27 | DeepSeekMath-V2 | weights/repository/report | based on V3.2-Exp-Base; added amount unknown [U] | repository, report |
DeepSeek-R1-Distill is not placed in this table a second time: all six students were one coordinated 2025-01-20 reasoning release and appear in the mainline table.
Vision-language, unified generation, and OCR¶
| Date | Model or event | Event type | Token/accounting status | Evidence |
|---|---|---|---|---|
| 2024-03-11 | DeepSeek-VL 1.3B / 7B | official family release | starts from about 500B / 2T text checkpoints; added joint total unknown | repository news, report |
| 2024-10-18 | Janus 1.3B | weights/repository | exact comparable total unknown [U] | registry, repository |
| 2024-11-13 | JanusFlow 1.3B | official repository news | exact comparable total unknown [U] | Janus news, report |
| 2024-12-13 | DeepSeek-VL2 Tiny / Small / 27B | official family release | about 818–831B across alignment, joint pretraining, and SFT depending on size | repository news, report |
| 2025-01-27 | Janus-Pro 1B / 7B | official repository news | training steps/mixtures disclosed; comparable token total unknown [U] | Janus news, report |
| 2025-10-20 | DeepSeek-OCR | official release | exact total unknown [U] | repository release log, report |
| 2026-01-27 | DeepSeek-OCR 2 | official release | exact total unknown [U] | OCR release log, OCR 2 repository, report |
These branches should not be forced into the V2 → V3 → V4 text-model chain. DeepSeek-VL and VL2 understand images; Janus adds image generation; OCR models specialize in document compression and recognition. Shared language backbones do not make their training exposures directly comparable.
Product aliases and names that are not separate generations¶
| Name | Correct interpretation at the cutoff |
|---|---|
deepseek-chat |
backward-compatible API alias whose backing checkpoint changes; the alias itself is not one stable model generation |
deepseek-reasoner |
reasoning-mode API alias; it has pointed to R1/V3-family and later hybrid-thinking behavior depending on release date |
| DeepSeek-R1-Lite-Preview | web-only preview that preceded R1; no open checkpoint or exact training-token total was disclosed |
| DeepSeek-V4-Pro-Max | maximum reasoning-effort mode of V4-Pro, not a separately disclosed checkpoint or pretraining run |
| V3.2-Speciale | temporary high-compute reasoning variant, not a new base family |
| R1-Distill-Qwen/Llama | six fine-tuned student checkpoints released together; their foundation pretraining belongs to Qwen/Llama, not DeepSeek |
Base, Chat, Instruct, RL, Zero |
training-stage or behavior suffixes; only count as a separate generation when the source reports a distinct checkpoint and recipe |
BF16, FP8, INT4, AWQ, GPTQ, GGUF |
precision or packaging variants; a conversion is not a new trained model unless the vendor reports additional training |
The DeepSeek change log should be read as a dated routing ledger. Calling
deepseek-chat in May 2024 and April 2026 does not imply that the same weights
served both requests.
Condensed chronology¶
2023 Coder v1 -- DeepSeek-LLM
2024 MoE -- Coder-v1.5 -- Math -- VL -- V2/Lite -- Prover --
Coder-V2 -- Prover-V1.5 -- V2.5 -- Janus -- JanusFlow --
VL2 -- V3
2025 R1/R1-Zero/Distill -- Janus-Pro -- V3-0324 -- Prover-V2 --
R1-0528 -- V3.1 -- Terminus -- V3.2-Exp -- OCR --
Math-V2 -- V3.2/Speciale
2026 OCR 2 -- V4-Flash/V4-Pro
Remaining unknowns¶
- Exact first-party announcement dates for some early weights whose official repositories provide no dated news item.
- Added training exposure for V2.5, dated V2/V3/R1 updates, production V3.2, and most specialist post-training branches.
- Unique-token counts, complete corpus identities and licenses, and sampling repetition for every flagship.
- Whether all live product surfaces route identically at every moment; API aliases can change without creating a new open checkpoint.
- Full pretraining, ablation, failed-run, and post-training compute for most generations.
An unknown field is intentionally retained rather than reconstructed from model names, benchmark behavior, or community packaging.