Short answer
DeepSeek-V3.2-Exp was an experimental model built on V3.1-Terminus. DeepSeek later described V3.2 as its official successor; Exp is not another name for V3.2-Speciale. For production API use, distinguish that release history from current routing: DeepSeek’s current pricing documentation lists deepseek-flash and deepseek-v4-pro, while its model-list endpoint documents how to query currently available models. The documentation reviewed here does not establish whether V3.2-Exp remains selectable through that endpoint, so verify the exact model ID and routing before deployment. (api-docs.deepseek.com)
What “DeepSeek 3.2 Exp” means
DeepSeek introduced V3.2-Exp in 2025 as an experimental model built on V3.1-Terminus, with DeepSeek Sparse Attention (DSA). DeepSeek later called V3.2 the official successor to V3.2-Exp and announced V3.2-Speciale as a separate model. These names and releases should not be treated as interchangeable. (api-docs.deepseek.com)
The current API documentation reviewed for this article identifies deepseek-flash and deepseek-v4-pro in its pricing guide. It says deepseek-flash is served by DeepSeek-V4.1-Flash-1017. Its model-list documentation describes an endpoint for listing currently available models, but the retrieved documentation does not provide a live response confirming whether V3.2-Exp is selectable. Treat the listed choices as documented API options—not as proof that Exp has been retired or is unavailable everywhere. (api-docs.deepseek.com)
Capabilities and benchmark trade-offs
SiliconFlow’s October 14, 2025 announcement promoted V3.2-Exp for long-context reasoning, coding, and agent-search tasks, and reported a 164K context window. This is a dated provider claim; the announcement does not clarify what the figure counts or establish the usable limit across providers. Confirm the context and output limits for the exact model and service you plan to use.
The announcement’s comparison with V3.1-Terminus shows a mix of numerical increases, decreases, and a tie. Selected scores as printed in its table:
| Benchmark | V3.1-Terminus | V3.2-Exp | Numerical comparison |
|---|---|---|---|
| BrowseComp | 38.5 | 40.1 | Higher for Exp |
| Codeforces-Div1 | 2046 | 2121 | Higher for Exp |
| AIME 2025 | 88.4 | 89.3 | Higher for Exp |
| GPQA-Diamond | 80.7 | 79.9 | Lower for Exp |
| Humanity’s Last Exam | 21.7 | 19.8 | Lower for Exp |
| LiveCodeBench | 74.9 | 74.1 | Lower for Exp |
| MMLU-Pro | 85.0 | 85.0 | Tied |
These are provider-reported figures, not an independent evaluation. The announcement does not supply enough methodology or metric detail to treat every numerical change as a practical quality improvement. Consider Exp for a controlled research or historical comparison; for a model choice, test representative tasks, prompts, and tool workflows on the exact service you would use.
Production status: model versus deployment
An experimental model and an unsupported deployment artifact are different things. NVIDIA’s NGC page displays an End of Support warning for its specific V3.2-Exp container. That warning applies to that artifact; it does not establish that every hosted API or self-hosted copy of the model is unavailable. The page’s separate version and tag metadata do not explain away the warning. Check the current status of any particular artifact before relying on it in production.
SiliconFlow’s 2025 announcement described Playground and API access at that time. It does not establish current availability or production service terms. Before committing to a host, confirm the exact model version and ID, endpoint, limits, tool behavior, pricing, data handling, and support terms with that provider.
API access and cost
SiliconFlow’s October 14, 2025 article showed a provider-specific example using POST https://api.siliconflow.com/v1/chat/completions, model ID deepseek-ai/DeepSeek-V3.2-Exp, and Bearer-token authorization. Treat it as a dated integration reference, not confirmation that the endpoint or ID remains active. The captured article does not provide a complete response schema or production error-handling guidance.
The same article listed $0.27 per million input tokens and $0.41 per million output tokens. Those were SiliconFlow’s stated rates in that 2025 announcement, not a current quote. Its “50% lower cost” headline did not identify a comparison model or calculation in the article. Check the provider’s current prices and billing terms before estimating spend.
DeepSeek’s pricing guide, reviewed November 10, 2026, lists the following prices per 1 million tokens for its named API models. It specifies peak hours of 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours, including weekends and Chinese public holidays, are off-peak. The guide notes that prices may change. These entries describe the models named in the pricing guide; they do not establish a current price for V3.2-Exp. (api-docs.deepseek.com)
| API model | Input, cache hit (off-peak / peak) | Input, cache miss (off-peak / peak) | Output (off-peak / peak) |
|---|---|---|---|
deepseek-flash (DeepSeek-V4.1-Flash-1017) |
$0.003 / $0.006 | $0.15 / $0.30 | $0.60 / $1.20 |
deepseek-v4-pro |
$0.022 / $0.044 | $0.66 / $1.32 | $1.98 / $3.96 |
USD per million tokens, as listed in DeepSeek’s pricing documentation. Check the live pricing page before budgeting. (api-docs.deepseek.com)
Practical selection checks
- Using DeepSeek’s API? Consult its current model list and pricing guide, then confirm that the exact model ID you intend to send is available. Do not assume the historical Exp identifier selects the same model today. (
api-docs.deepseek.com) - Evaluating Exp for research? Verify the weights, license, inference instructions, and the support status of your chosen serving framework or deployment artifact.
- Considering a third-party host? Confirm model version, endpoint, context and output limits, billing basis, and support terms directly with that provider.
- Comparing quality or cost? Run a representative pilot and measure task success, latency, retries, tool behavior, and cost per completed task—not only price per million tokens.
Bottom line: V3.2-Exp was an experimental predecessor to V3.2, not a synonym for V3.2 or V3.2-Speciale. Its reported benchmark results are mixed, and its current API availability is not established by the documentation reviewed here. Treat it as an evaluation target unless the provider confirms access and terms for the exact Exp model you need.
Sources
DeepSeek-V3.2 Release — DeepSeek API Docs
Lists Models — DeepSeek API Docs
Models & Pricing — DeepSeek API Docs