DeepSeek-V3.2

What it is. DeepSeek-V3.2 (released 2025-12-01, MIT open weights) is the previous DeepSeek flagship: a DeepSeek Sparse Attention (DSA) MoE that introduced the sparse-attention direction the V4 family later commercialized. Why it matters. It is the MIT-licensed checkpoint that marked DeepSeek’s turn from dense attention toward sparse-attention cost economics, and it remains a downloadable artifact for self-hosters and researchers even though the API is retired. Key characteristics. 61 layers, hidden size 7168, 256 routed experts (8 selected + 1 shared), vocab 129,280 (from config.json); reasoning, coding, tool-calling and agent capabilities. What a professional should know. It is discontinued on the DeepSeek API (superseded by the V4 family), the release page carries qualitative rather than numeric performance claims, and no numeric context window or benchmark scores were published.

Status and successor

DeepSeek-V3.2 is discontinued — superseded by DeepSeek-V4-Pro and the wider V4 family on the DeepSeek API. It is kept in the database for historical and open-weight reference. Because it is API-discontinued, no current API pricing applies. China AI Hub analysis indicates V3.2’s supersession is the pivot point where DeepSeek moved from a conventional MoE (V3.2’s 256-expert dense-attention design) to the sparse-attention V4 generation, so the model’s historical role is as the last of the pre-sparse DeepSeek flagships.

Architecture and parameters

The architecture is recorded from the Hugging Face config.json: 61 layers, hidden size 7168, MoE intermediate 2048, 256 routed experts with 8 selected + 1 shared per token, and a 129,280 vocab. DeepSeek Sparse Attention (DSA) is the headline innovation — a linear-attention-style mechanism that makes long context cheaper by cutting the quadratic cost of standard attention. China AI Hub analysis indicates DSA is the architectural through-line from V3.2 to the V4 family: V3.2 introduced the direction, and V4.1-Flash’s 552B asymmetric sparse design is the commercial maturation of the same cost-per-token logic. See Mixture-of-Experts and the MoE architectures research.

What is and is not documented

DeepSeek’s own release pages do not state a numeric context window or benchmark scores — the release page carries qualitative claims (“GPT-5 level performance”) and notes that the API-only V3.2-Speciale variant rivaled Gemini-3.0-Pro with gold-medal results in math/programming olympiads, but lists no numbers. China AI Hub analysis indicates the absence of numeric scores on a flagship release is itself informative: it signals a positioning-by-claim rather than positioning-by-benchmark release, which the V4 family later reversed by publishing concrete benchmark tables.

Open weights and license

MIT-licensed open weight and self-hostable. This is the most permissive license in the collection, and — combined with the fact that the weights remain downloadable after API discontinuation — it is the main reason V3.2 retains reference value. The practical caveat is that this is a 2025-generation model: a 256-expert dense-attention MoE of this scale is a substantial multi-GPU deployment, and it is no longer a frontier model. See licensing explained and open weight vs API.

Benchmark interpretation

There are no numeric benchmark rows recorded: the release page publishes qualitative comparisons only. This is a genuine absence in the source material, not a gap in the database — DeepSeek did not publish a numeric score table for V3.2 on the fetched release pages.

What the benchmarks do not prove

With no numeric scores, there is nothing with which to rank V3.2 against competitors, and the “GPT-5 level performance” and “gold-medal olympiad” claims are qualitative and cannot be independently verified from the release page. Treat capability as vendor-positioned. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: historical and open-weight research, self-hosting where a permissive MIT license and a downloadable 2025-generation checkpoint are the goal, and reproducibility work on the sparse-attention lineage. Less suited: any production API workload (discontinued), frontier-scale reasoning (superseded), and long-context work without a documented context window.

Where this model fits

Workload	Relevance
Long-context analysis	Unknown
Coding	Moderate
Structured API workflows	No evidence
Agent orchestration	Unknown
Local self-hosted deployment	High
GUI automation	No evidence
Video generation	No evidence
Enterprise cloud	No evidence

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims. API-dependent rows are No evidence because the model is API-discontinued; long-context and agent rows are Unknown because no context window or benchmark data was published.

Historical significance

China AI Hub analysis indicates V3.2’s enduring relevance is as the architectural waypoint between DeepSeek’s dense-attention past and its sparse-attention future: it introduced DSA and the MIT-licensed checkpoint that preceded the V4 generation, and its qualitative-only release framing is the contrast against which the V4 family’s benchmark-table releases should be read. For self-hosters and researchers it is a downloadable artifact; for API users it is retired. See how DeepSeek changed China’s AI market for the market context.

What is uncertain

• The numeric context window is not stated on the official release pages fetched.
• Vendor performance claims are qualitative; no numeric benchmark scores were published.
• Whether the MIT-licensed weights match the final API checkpoint is not separately documented.

Sources

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-models-deepseek-v3-2-1	DeepSeek-V3.2 release	https://www.deepseek.com/en/news/deepseek-v3-2/	Official	2025-12-01	2026-09-20	high	—
src-models-deepseek-v3-2-2	DeepSeek Transparency Center	https://www.deepseek.com/en/transparency/	Official	—	2026-09-20	high	—
src-models-deepseek-v3-2-3	Hugging Face model card — DeepSeek-V3.2	https://huggingface.co/deepseek-ai/DeepSeek-V3.2	Model card	—	2026-09-20	high	—
src-models-deepseek-v3-2-4	Hugging Face config.json — DeepSeek-V3.2 (layer/expert/hidden dimensions)	https://huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/main/config.json	Official documentation	—	2026-09-27	high	—

Labels used above: Official fact (architecture from config.json, release facts), Vendor-reported claim (the “GPT-5 level performance” claim), and China AI Hub analysis (our synthesis, introduced as such). Context window and benchmark scores are not publicly documented for this model.