Short answer: DeepSeek-V3.2-Exp was an experimental model built on V3.1-Terminus to explore DeepSeek Sparse Attention (DSA) and improve efficiency on long-context tasks. DeepSeek later named V3.2 the official successor. The current API model documentation reviewed here lists deepseek-flash and deepseek-v4-pro, not V3.2-Exp, so it does not establish a current official API identifier for calling the exact experimental model. If you can confirm access to that version through another channel, it may be worth testing on long-input workloads—but the published results do not prove it is better for your particular application. (DeepSeek’s V3.2-Exp announcement; V3.2 announcement; current API quickstart)
Official documentation reviewed November 8, 2026. Model availability and prices can change.
What was DeepSeek-V3.2-Exp designed to do?
DeepSeek described V3.2-Exp as an experimental intermediate step built on V3.1-Terminus. Its model card says DSA uses fine-grained sparse attention and was designed to explore training and inference efficiency in long-context scenarios. That makes long documents or multi-document inputs reasonable workloads to evaluate—not proven cases where the model will outperform another option. (DeepSeek-V3.2-Exp model card)
DeepSeek published benchmark comparisons with V3.1-Terminus. In the model card’s reasoning mode without tool use results, V3.2-Exp scored 79.9 versus 80.7 on GPQA-Diamond, 74.1 versus 74.9 on LiveCodeBench, and 89.3 versus 88.4 on AIME 2025. These are DeepSeek-reported scores for those evaluations; they do not measure your workload’s latency, cost, or accuracy. The same table separately reports agentic tool-use results, so the scores should not be read as one uniform evaluation setting. (Model card and benchmark table)
DeepSeek’s launch announcement also said API prices were cut by more than 50% at that time. That is a historical launch-period statement, not a current V3.2-Exp rate or a basis for a present-day budget. (Launch announcement)
Can you call V3.2-Exp through DeepSeek’s production API now?
The December 1, 2025 announcement calls DeepSeek-V3.2 the official successor to V3.2-Exp and says V3.2 was available through the App, Web, and API. The current API quickstart and pricing documentation list deepseek-flash and deepseek-v4-pro; they do not list V3.2-Exp. Based on those public model listings, a current official API identifier for the exact experimental model is not established. This is distinct from whether a third-party host or local deployment offers the weights. (V3.2 announcement; API quickstart; model and pricing list)
Do not rely on older examples using model="deepseek-chat" to identify V3.2-Exp. DeepSeek’s changelog said the legacy deepseek-chat and deepseek-reasoner names were scheduled for discontinuation on July 24, 2026; the current quickstart instead names the current model identifiers. The old alias does not establish access to a particular V3.2-Exp version. (DeepSeek changelog; current API quickstart)
Example for a currently listed API model—not V3.2-Exp
DeepSeek’s current quickstart documents an OpenAI-compatible API format. The example below targets deepseek-flash, not V3.2-Exp; current pricing documentation identifies the served model version as DeepSeek-V4.1-Flash-1017. (Quickstart; model and pricing list)
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DEEPSEEK_API_KEY"],
base_url="https://api.deepseek.com",
)
response = client.chat.completions.create(
model="deepseek-flash",
messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)
This is a current-model example, not a version-pinned V3.2-Exp request. For a new integration, use the identifier and request format in the current official docs, then test the actual behavior you need before deployment. (Chat Completions API reference)
Current API pricing for a practical alternative
DeepSeek’s current pricing page lists rates per 1 million tokens, separating input cache hits, input cache misses, and output. These rates apply to deepseek-flash, not V3.2-Exp. The page displays dollar signs but does not spell out a currency code in the retrieved pricing text, so confirm the billing currency and current cache-miss rate on the official page before budgeting. (DeepSeek pricing documentation)
DeepSeek defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays; other times are off-peak. The pricing page says off-peak rates are half of peak rates and advises checking the page for current prices. (Pricing and peak-hour conditions)
When is V3.2-Exp worth evaluating?
Consider testing V3.2-Exp if you have confirmed access to that exact version and your workload has long inputs—for example, analyzing a large report or synthesizing several documents. Those are test scenarios suggested by DSA’s stated long-context efficiency goal, not guaranteed strengths. Compare it with an available alternative using the same representative inputs, and measure task accuracy, latency, and the cost for the provider or deployment channel you would actually use.
If you need a newly documented hosted API integration rather than the experimental version specifically, start with the model identifiers currently listed in DeepSeek’s API documentation. Do not assume that a generic endpoint or legacy alias selects V3.2-Exp.
Bottom line
V3.2-Exp was an experimental long-context efficiency step, and DeepSeek later identified V3.2 as its official successor. DeepSeek’s benchmark table provides comparisons with V3.1-Terminus, but those scores alone cannot predict results on your workload. The current API model listings reviewed here do not provide a V3.2-Exp identifier. Test the exact model only after confirming access; for a new official API integration, use a currently documented model and compare quality, latency, and current token costs on your own inputs.
Sources
Introducing DeepSeek-V3.2-Exp | DeepSeek API Docs
DeepSeek-V3.2 Release | DeepSeek API Docs
Your First API Call | DeepSeek API Docs
Models & Pricing | DeepSeek API Docs