DeepSeek-V3.2

DeepSeek-V3.2 is an open-weights model from the DeepSeek family that is positioned around efficient reasoning and agent-style use, pairing a new sparse attention design with a heavily scaled post-training reinforcement learning stage. According to the official model card on Hugging Face, the release introduces DeepSeek Sparse Attention (DSA), which lowers the computational cost of attention specifically for long-context scenarios while preserving overall model quality. The same card highlights a scalable RL framework in which post-training compute is increased substantially so that DeepSeek-V3.2 reaches reasoning performance comparable to GPT-5, framing the model as a deliberate balance between efficiency and capability for agentic AI applications.

Independent reporting describes DeepSeek-V3.2 as a family of open-source reasoning models, with a higher-compute variant called DeepSeek-V3.2-Speciale that is reported to outperform GPT-5 and match Gemini-3.0-Pro on reasoning benchmarks. Community coverage notes that, in practice, the model offers solid reasoning quality and unusually low cost per token, though generation throughput can be slower than competing frontier systems, making it a reasonable fit for workloads that value reasoning depth, long-context handling, and open-weight deployment over raw latency.

Vivgriddeepseek-v3.2deepseek

Open provider page Provider docs

Quick Info

Provider: Vivgrid
Model key: deepseek-v3.2
Release date: Dec 1, 2025
Last updated: Dec 1, 2025
Knowledge cutoff: 2024-07
AI SDK package: @ai-sdk/openai-compatible
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.28
Output token cost: $0.42

Limits

Output tokens: 128,000 tokens
Context window: 128,000 tokens

Transparent token rates

Compare DeepSeek-V3.2 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about DeepSeek-V3.2

Vivgrid

CoveragePreview

DeepSeek says both models are more efficient and performant than DeepSeek V3.2 due to architectural improvements, and have almost "closed the gap" with current leading models, both open and closed, on reasoning benchmarks.

Vivgrid

Coverage

DeepSeek's official changelog documents that on 2025-12-01, both the deepseek-chat and deepseek-reasoner API aliases were upgraded to DeepSeek-V3.2, with deepseek-chat mapping to non-thinking mode and deepseek-reasoner to thinking mode, and DeepSeek-V3.2-Speciale served via a temporary endpoint. This confirms V3.2 as t The same changelog records the 2026-04-24 launch of DeepSeek V4 (deepseek-v4-pro and deepseek-v4-flash) on both OpenAI ChatCompletions and Anthropic interfaces, with the legacy deepseek-chat and deepseek-reasoner aliases slated for discontinuation on 2026-07-24 and repointed to V4-Flash modes during the transition. A 2

Vivgrid

CoverageBenchmark

DeepSeek V3.2-Speciale achieves 96% on AIME, gold at IMO, and top-10 at IOI—matching U.S. frontier models despite export restrictions.

Vivgrid

Coverage

DeepSeek released DeepSeek-V3.2, a family of open-source reasoning and agentic AI models. The high compute version, DeepSeek-V3.2-Speciale, performs better than GPT-5 and comparably to Gemini-3.0-Pro

Vivgrid

CoverageBenchmark

A deep technical breakdown of DeepSeek V3.2, examining how training data, synthetic pipelines, sparse attention, and post-training RL shape reasoning and performance.

Vivgrid

Coverage

Updated March 2026: Comprehensive guide to DeepSeek V3.2, V4 (expected April 2026), R1/R2 reasoning models, and how to use DeepSeek in Antigravity via the OpenAI compatibility layer.