DeepSeek V3.2 explained: sparse attention, benchmarks and what replaced it

If you shipped anything against DeepSeek’s API in late 2025, you almost certainly ran into DeepSeek V3.2. Its experimental predecessor, V3.2-Exp, arrived on 29 September 2025 with DeepSeek Sparse Attention (DSA) and an API price cut of more than 50%; the official V3.2 followed on 1 December 2025 as its successor. On 24 April 2026 the DeepSeek-V4 preview reached the API and the lineup changed again — today DeepSeek’s API serves deepseek-flash and deepseek-v4-pro. This article covers what V3.2 actually was (architecturally, not in marketing terms), how it scored against V3.1-Terminus, what it cost on the API, and what changes if you are moving an old integration to today’s models.

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 on 1 December 2025, with code and weights on Hugging Face under the MIT License. DeepSeek’s release note calls it the “official successor to V3.2-Exp”, and it was the model behind DeepSeek’s API until the V4 preview launched on 24 April 2026. The headline architectural change versus V3.1-Terminus is a single one: per the V3.2 technical report, “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 through continued training.

V3.2 is now a previous generation. According to DeepSeek’s API Change Log, deepseek-chat and deepseek-reasoner were upgraded to V3.2’s non-thinking and thinking modes on 1 December 2025; on 24 April 2026 both names were pointed to deepseek-v4-flash (itself now a legacy name served by DeepSeek-V4.1-Flash-1017), with retirement scheduled for 2026-07-24 15:59 UTC, and they are no longer listed in DeepSeek’s docs. The API now offers deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). See the full lineup on the DeepSeek models hub.

Architecture and lineage

V3.2 was built by continued training from a DeepSeek-V3.1-Terminus checkpoint, so it keeps the MoE backbone that ran from V3 through V3.1 — listed on the V3.1 model card as 671B total parameters with 37B activated per token and a 128K context. 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. Per the V3.2 report, 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 key-value entries with the top-k index scores feed the main attention step (the official inference config sets k to 2,048). DSA reduces the main model’s core attention complexity from quadratic O(L²) to O(Lk), where k (≪ L) is the number of selected tokens; the indexer itself is still O(L²) but, the report says, needs far less computation than the MLA attention in V3.1-Terminus.

DeepSeek kept the training recipe deliberately conservative to isolate the architectural effect. The V3.2-Exp model card says DSA delivers “substantial improvements in long-context training and inference efficiency while maintaining virtually identical model output quality”, and that its training configurations were deliberately aligned with V3.1-Terminus to evaluate the impact. V3.2 then added a scaled reinforcement-learning post-training phase — the report puts the post-training compute budget at more than 10% of pre-training cost — on top of the same architecture.

Variants shipped under the V3.2 name

• DeepSeek-V3.2 — the general model served on deepseek-chat (non-thinking) and deepseek-reasoner (thinking) from 1 December 2025 until 24 April 2026.
• DeepSeek-V3.2-Exp — the 29 September 2025 experimental predecessor, which DeepSeek described as an intermediate step toward its next-generation architecture, built to validate DSA’s efficiency gains.
• DeepSeek-V3.2-Speciale — a high-compute reasoning variant. API-only at launch (its weights were also published on Hugging Face), it was served via a temporary endpoint until 15 December 2025, 15:59 UTC, at the same price as V3.2 and without tool calls, “to support community evaluation & research”.

Context, licensing and release timeline

For the older base model V3.2 descends from, see DeepSeek V3. For the models that 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. The numbers below are DeepSeek-reported, from the V3.2-Exp model card (September 2025; SWE-Bench Verified and BrowseComp are from the card’s agentic tool-use section, the others from reasoning mode without tool use).

The December V3.2 release then pushed further. The V3.2 technical report (arXiv 2512.02556) says V3.2 “performs comparably to GPT-5”, and that the high-compute V3.2-Speciale “surpasses GPT-5” and reasons on par with Gemini-3.0-Pro, with gold-medal performance in IMO 2025 and IOI 2025; DeepSeek’s release note adds gold-level results at CMO and the ICPC World Finals. Selected numbers from the report’s tables (all DeepSeek-reported, thinking mode):

The report also notes that Speciale reaches these scores with longer outputs than V3.2, and that V3.2 itself still trails frontier closed models such as Gemini-3.0-Pro on breadth of world knowledge and token efficiency.

Strengths — where V3.2 actually won

• Long-context economics. DSA’s O(Lk) core attention is the whole point. The V3.2 report estimates per-token costs from DeepSeek’s H800 deployment and shows V3.2’s cost rising far more slowly with token position than V3.1-Terminus’s.
• Agentic tool-use. DeepSeek’s 1 December 2025 release note calls V3.2 its “first model to integrate thinking directly into tool-use”, with tool calls supported in both thinking and non-thinking modes. That combination made it a credible open choice for agent frameworks.
• Price. The V3.2-Exp release on 29 September 2025 came with an API price cut of more than 50%, and the December V3.2 release kept those rates (see the pricing snapshot below).
• Open weights. MIT-licensed weights meant you could self-host the same model DeepSeek served on its 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 today’s deepseek-flash and deepseek-v4-pro offer a 1M-token context with up to 384K output tokens. The V3.2 report itself notes that roughly 20% or more of its search-agent test cases exceeded the 128K limit.
• No image input. V3.2 is a text model; for vision you needed DeepSeek VL2. (On today’s API, deepseek-flash accepts images.)
• Custom inference kernels. DSA is not a drop-in attention backend. DeepSeek released new kernels alongside the model — TileLang reference kernels, indexer kernels in DeepGEMM and sparse-attention kernels in FlashMLA — and on 17 November 2025 it flagged a RoPE layout discrepancy in the indexer of its earlier inference demo code that could degrade performance, so self-hosters on the old demo code had to update.
• Speciale’s tool-use gap. The Speciale reasoning variant did not support tool calls 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, but DeepSeek no longer serves it: it is not on the current pricing page, and the official web and app models were updated to DeepSeek-V4.1-Flash-1017 on 17 October 2026, per the Change Log. Your options:

1. Official API (current models only). 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. V3.2-era code that called deepseek-chat or deepseek-reasoner needs to move to deepseek-flash or deepseek-v4-pro; DeepSeek’s own guidance is to “change the model name to deepseek-flash” and keep the base URL. See DeepSeek API documentation.
2. Self-host. Download the weights (about 689.5 GB of files) from Hugging Face and serve with vLLM or SGLang. V3.2 shares V3.2-Exp’s model structure, and the V3.2-Exp card gives the SGLang launch command python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --enable-dp-attention; vLLM provided day-0 support for V3.2-Exp. For local use, the V3.2 card recommends temperature = 1.0, top_p = 0.95. Hardware planning matters — start at the DeepSeek hardware calculator.

Pricing snapshot

V3.2’s API rates were $0.028 per million input tokens on a cache hit, $0.28 on a cache miss and $0.42 per million output tokens, a single rate at all hours (DeepSeek’s pricing page listed no off-peak tier in that period). DeepSeek’s V3.2-Exp release note gives these rates as effective from 10:00 UTC on 29 September 2025, down from $0.07 / $0.56 / $1.68, and Wayback Machine captures of DeepSeek’s pricing page (11 October 2025, 12 January and 11 April 2026) show them unchanged through the V3.2 period. These are historical rates; V3.2 is no longer on the pricing page.

Today’s deepseek-flash does not undercut V3.2 across the board. As of October 2026 it costs $0.006 (peak) / $0.003 (off-peak) per million cache-hit input tokens, $0.30 / $0.15 per million cache-miss input tokens and $1.20 / $0.60 per million output tokens; deepseek-v4-pro costs $0.044 / $0.022, $1.32 / $0.66 and $3.96 / $1.98. Off-peak rates are half of peak; peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. Check the DeepSeek API pricing guide and the DeepSeek pricing calculator before you quote costs to a stakeholder.

Worked example — V3.2’s old rates vs deepseek-flash today

1,000,000 API calls with a 2,000-token cached system prompt, a 200-token uncached user message and a 300-token response — 2,000M cache-hit input tokens, 200M cache-miss input tokens and 300M output tokens:

For this output-heavy mix, deepseek-flash at peak costs more than V3.2 did, because its output rate is higher; off-peak it comes in slightly below. 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 are picking a model today, these are the workloads where V3.2’s open weights (or today’s deepseek-flash and deepseek-v4-pro) fit:

• 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

DeepSeek’s own V3.2 report compared it with Claude-4.5-Sonnet, GPT-5 High and Gemini-3.0-Pro among closed models, and with Kimi-K2-Thinking and MiniMax M2 among open models that support thinking in tool use. For structured comparisons: DeepSeek vs ChatGPT, DeepSeek vs Claude, and the broader DeepSeek alternatives roundup.

Verdict

DeepSeek V3.2 did what DeepSeek said it would: it kept V3.1-Terminus quality on DeepSeek’s benchmarks, brought sparse attention and a price cut of more than half to the API, and seeded the agentic tool-use and scaled-RL patterns that later models built on. It was the model behind DeepSeek’s API from December 2025 until April 2026 (with V3.2-Exp from late September). It is not the model to start a new integration on today — DeepSeek no longer serves it, so move hosted workloads to deepseek-flash or deepseek-v4-pro, keeping the same base_url and changing only the model field, or self-host the MIT weights if you specifically need V3.2.