If you shipped anything against DeepSeek’s API in late 2025, you almost certainly ran into DeepSeek V3.2. It landed on December 1, 2025 as the official successor to V3.2-Exp, carried an aggressive price drop, and introduced DeepSeek Sparse Attention (DSA) — the architectural change that made long-context calls roughly half the cost overnight. Four and a half months later, on April 24, 2026, DeepSeek V4 arrived and reshuffled the model lineup again. This article covers what V3.2 actually was (architecturally, not in marketing terms), how it benchmarked against V3.1-Terminus, what it cost to run, and what changes in practice if you are migrating an integration forward to V4-Flash or V4-Pro today.

What DeepSeek V3.2 is (and where it sits now)

DeepSeek V3.2 is an open-weight mixture-of-experts (MoE) language model released by DeepSeek-AI on December 1, 2025, distributed on Hugging Face under the MIT License. It is the official, production-ready successor to the September experimental release V3.2-Exp, and it was the company’s flagship chat model until DeepSeek V4 launched on April 24, 2026. The headline architectural change versus V3.1-Terminus is a single one: DeepSeek-V3.2 uses exactly the same architecture as DeepSeek-V3.2-Exp, and compared with V3.1-Terminus the only architectural modification is the introduction of DeepSeek Sparse Attention (DSA) through continued training.

As of April 24, 2026, V3.2 is the previous generation. The legacy API IDs deepseek-chat and deepseek-reasoner, which pointed at V3.2 until the V4 release, now route to deepseek-v4-flash and will be retired at 2026-07-24 15:59 UTC. Any production integration still calling those IDs should plan a one-line swap within that window. See the full lineup on the DeepSeek models hub.

Architecture and lineage

V3.2 inherits the 671B-total-parameter MoE backbone introduced with V3, continued through V3.1 and V3.1-Terminus. What is new is how attention is computed.

DeepSeek Sparse Attention (DSA)

Classical transformer attention is quadratic: every token attends to every other token, so doubling the sequence length quadruples the compute. The DSA prototype consists of two components — a lightning indexer and a fine-grained token selection mechanism. The lightning indexer computes an index score between the current query token and each preceding token, determining which tokens the query should attend to. Only the top-k selected key-value entries feed the main attention step. DSA reduces attention’s computational complexity from quadratic O(L²) to linear O(Lk), where k (≪ L) is the number of selected tokens.

DeepSeek kept the training recipe deliberately conservative to isolate the architectural effect. DSA achieves fine-grained sparse attention while maintaining virtually identical model output quality, and the training configurations of V3.2-Exp were deliberately aligned with V3.1-Terminus to evaluate the impact cleanly. V3.2 then added a scaled reinforcement-learning post-training phase on top of the same architecture.

Variants shipped under the V3.2 name

• DeepSeek-V3.2 — the daily-driver model served on deepseek-chat (non-thinking) and deepseek-reasoner (thinking) during the V3.2 window.
• DeepSeek-V3.2-Exp — the September 29, 2025 experimental predecessor, shipped first to stabilise serving infrastructure.
• DeepSeek-V3.2-Speciale — a high-compute reasoning variant. API-only, served via a temporary endpoint and intentionally without tool-calling to support community evaluation. It targeted maximum reasoning effort rather than everyday chat.

Context, licensing and release timeline

For the older base model V3.2 was trained from, see DeepSeek V3. For the next generation that has replaced it, see DeepSeek V4.

Benchmarks: V3.2 vs V3.1-Terminus and the Speciale headline

DeepSeek designed V3.2-Exp as an architectural ablation: the point was to show parity with V3.1-Terminus while the attention cost dropped, not to claim new highs. Numbers below are from DeepSeek’s own report and the llm-stats summary of the September 2025 release; cross-check the V3.2 technical report on arXiv if you need the exact table.

The V3.2 release (December 2025) then pushed further. DeepSeek-V3.2 harmonises high computational efficiency with stronger reasoning and agent performance, and the company reported it performs comparably to GPT-5 on its scalable RL framework. The extreme variant, V3.2-Speciale, was positioned specifically for long reasoning: Speciale achieves performance parity with Gemini-3.0-Pro and showed gold-medal performance in IOI 2025, ICPC World Final 2025, IMO 2025 and CMO 2025. Treat those as vendor-reported; external leaderboards typically show narrower gaps.

Strengths — where V3.2 actually won

• Long-context economics. DSA’s O(Lk) scaling is the whole point. At 128K tokens, V3.2 processed inputs far more cheaply than any dense-attention peer with the same parameter count.
• Agentic tool-use. V3.2 was DeepSeek’s first model to integrate thinking directly into tool-use, supporting tool-calling in both thinking and non-thinking modes. That combination made it a credible open choice for agent frameworks.
• Price. The September V3.2-Exp release triggered an API price cut of more than 50%, and the December V3.2 release kept those rates.
• Open weights. MIT-licensed weights meant you could self-host the same model used on the official API, which matters for regulated or air-gapped environments. See is DeepSeek open source for the licensing detail per release.

Weaknesses — where it fell short

• Context ceiling. 128K was generous in late 2025 but is now outclassed by V4’s 1,000,000-token default.
• No native multimodal input. V3.2 is a text model; for vision you needed DeepSeek VL2.
• Custom inference kernels. DSA is not a drop-in attention backend. Running weights locally required the FlashMLA / DeepGEMM / TileLang kernels DeepSeek released alongside the model, and early inference code had a subtle RoPE bug that was fixed on 2025-11-17 — self-hosters who pulled weights before then had to re-test.
• Speciale’s tool-use gap. The Speciale reasoning variant didn’t support tool-calling during its temporary endpoint window, so agentic workloads had to stay on standard V3.2.

How to access V3.2 today

DeepSeek V3.2 remains available as open weights and is still referenced on DeepSeek’s pricing documentation during the V4 Preview window, but the default chat and API experience has moved to V4. Your three paths:

1. Official API. Chat requests hit POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. The API is stateless — clients must resend the entire conversation history on every request, which is a different contract from the web and mobile apps, which maintain session history for you. For new work, use the V4 model IDs directly; for legacy V3.2-era code, the old IDs still resolve (to V4-Flash now) until July 24, 2026. See DeepSeek API documentation.
2. Self-host. Download from Hugging Face and serve with vLLM or SGLang. SGLang launches with python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention, and vLLM provided day-0 support. Hardware planning matters — start at the DeepSeek hardware calculator.
3. Web and mobile apps. The apps now default to V4. V3.2 is no longer user-selectable there.

A minimal V4-Flash Python call using the OpenAI SDK looks like this:

from openai import OpenAI

client = OpenAI(base_url="https://api.deepseek.com", api_key="...")
resp = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "Summarise this contract in three bullets."}],
    temperature=1.0,
    max_tokens=1024,
)
print(resp.choices[0].message.content)

Swap model to deepseek-v4-pro for frontier-tier work, or add reasoning_effort="high" plus extra_body={"thinking": {"type": "enabled"}} to get reasoning_content alongside the final content.

Pricing snapshot

V3.2 pricing applied from September 29, 2025 (the V3.2-Exp drop) through the V4 release. Those rates — $0.28 per million input tokens (cache miss) and $0.42 per million output tokens — are no longer current. As of April 2026, V4-Flash has undercut them. Use the live DeepSeek API pricing page and the DeepSeek pricing calculator before you quote costs to a stakeholder.

Worked example — V4-Flash (what V3.2 workloads should migrate to)

1,000,000 API calls with a 2,000-token cached system prompt, a 200-token uncached user message, and a 300-token response, at V4-Flash rates ($0.028 cache-hit / $0.14 cache-miss / $0.28 output, per 1M tokens):

• Cached input: 2,000 × 1,000,000 = 2.0B tokens × $0.028/M = $56.00
• Uncached input: 200 × 1,000,000 = 0.2B tokens × $0.14/M = $28.00
• Output: 300 × 1,000,000 = 0.3B tokens × $0.28/M = $84.00
• Total: $168.00

Enumerating all three buckets matters — the user message on every call is still a cache miss against the cached system prefix, so leaving that line out understates real spend.

Best use cases

V3.2’s strengths mapped neatly to a handful of roles. If you’re picking a model today, these are the workloads where V3.2 (or its V4 successor) fits:

• Long-document analysis. Contracts, filings, codebases — see DeepSeek for legal research.
• Research assistance. Synthesising across papers and sources — see DeepSeek for research.
• Coding agents. Thinking-mode + tool-use is a workable agent loop — see DeepSeek for coding.
• RAG backends. Open weights + DSA favour RAG setups with large retrieved contexts — see the DeepSeek RAG tutorial.

Comparable alternatives

Anyone evaluating V3.2 in 2025 was typically also evaluating GPT-5-family models and Claude 4-family models. In the open-weights tier, the main competitors were Kimi K2 and Qwen-3-Max on coding; on reasoning, DeepSeek’s own R1 family was still the internal benchmark. For structured comparisons: DeepSeek vs ChatGPT, DeepSeek vs Claude, and the broader DeepSeek alternatives roundup.

Verdict

DeepSeek V3.2 did exactly what DeepSeek said it would: it kept V3.1-Terminus quality, halved long-context costs through DSA, and seeded the agentic tool-use and scaled-RL patterns that V4 then built on. It was the right model to run from October 2025 through April 2026. It is not the right model to start a new integration on today — migrate to V4-Flash for chat workloads or V4-Pro for frontier agentic work, keeping the same base_url and changing only the model field.

Last verified: 2026-04-24. DeepSeek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the official DeepSeek documentation and pricing page before committing to a production decision.