DeepSeek-V3.2 — reviews, specs & pricing

671B open-weight MoE with sparse attention for cheap long-context reasoning.

Summary

Open-weight successor to V3.2-Exp that introduces DeepSeek Sparse Attention for efficient long-context reasoning and agentic work. A high-compute Speciale variant competes with closed frontier models on reasoning benchmarks.

Sample use case

Open-source reasoning and agentic workloads at reduced inference cost.

Specifications

  • Provider: deepseek
  • License: open
  • Parameters: 671B
  • Context: 128k tokens
  • Released: 2025-12-01

Average rating 3.5 from 4 community reviews on Reviuws.

Community reviews

Simon Willison's blog on DeepSeek-V3.2

Rating: 4.0 / 5 — by Simon Willison (use case: general chat and agentic tasks)

Willison covers the simultaneous release of V3.2 and V3.2-Speciale as MIT-licensed 685B-parameter models, with V3.2 blending reasoning, agent and alignment data distilled from Speciale for day-to-day use.

Pros: Open MIT licence at frontier scale; agentic and reasoning capability distilled from a specialist model.

Cons: 690GB download puts self-hosting out of reach for most; confusing variant naming.

Simon Willison's blog on DeepSeek-V3.2

Rating: 3.0 / 5 — by Simon Willison (use case: long-context efficiency testing)

Testing the earlier V3.2-Exp release with his pelican-on-a-bicycle SVG benchmark, Willison describes an intermediate architecture step introducing sparse attention for long-context efficiency, with quality comparable to V3.1-Terminus.

Pros: New sparse attention targets long-context efficiency; open weights with a detailed tech report.

Cons: Explicitly experimental; no dramatic quality leap over V3.1.

VentureBeat on DeepSeek-V3.2

Rating: 4.0 / 5 — by Michael Nuñez (use case: cost-sensitive API deployment)

VentureBeat frames the release around affordability, reporting it mostly matches or slightly improves on V3.1's benchmarks while cutting API input pricing to under three cents per million tokens.

Pros: API pricing roughly halved; benchmarks hold steady or improve.

Cons: Gains are mostly cost rather than capability; experimental release.

The Neuron on DeepSeek-V3.2

Rating: 3.0 / 5 — by Grant Harvey (use case: cheap long-context work)

Harvey takes a sceptical but informative tone, asking whether the V3.2 pair is still a big deal in a crowded open-model market while acknowledging it meaningfully lowers the cost of long-context work.

Pros: Makes long-context inference notably cheaper; two variants to choose from.

Cons: Crowded market dulls the novelty; differentiation less clear-cut than before.

Frequently asked questions

What is DeepSeek-V3.2?

Open-weight successor to V3.2-Exp that introduces DeepSeek Sparse Attention for efficient long-context reasoning and agentic work. A high-compute Speciale variant competes with closed frontier models on reasoning benchmarks.

How much does DeepSeek-V3.2 cost?

Pricing for DeepSeek-V3.2 is not publicly listed.

Is DeepSeek-V3.2 open source?

DeepSeek-V3.2 is released under the open license.