DeepSeek V3.2 by NanoGPT | AI model information

$10 off the fastest Kimi K3 and GLM 5.2$10 off the fastest Kimi K3 and GLM 5.2

Model details

DeepSeek V3.2

DeepSeek V3.2 is positioned as a general-purpose flagship large language model built on DeepSeek Sparse Attention (DSA), an attention mechanism designed to keep compute and memory costs more manageable on long-context and agentic workloads. It ships alongside a sibling variant called V3.2-Speciale focused on reasoning, and both are released as open weights under the MIT license at roughly 685 billion parameters, giving developers freedom to self-host or use hosted APIs. A large-scale agentic reinforcement-learning pipeline underpins tool-use behavior, so the model is meant to be a drop-in workhorse for chat, retrieval, document analysis, and multi-step assistants rather than a narrow single-task system.

Through NanoGPT, V3.2 becomes practical for production deployments without managing infrastructure: the deployment accepts text and PDF input and returns text output, supports tool calling, structured output, and attachments, and exposes a 163,000-token context window that fits very long documents and extended agent sessions. Pricing is competitive for an open-weight flagship, with cache reads priced below standard input and output, which makes large-context retrieval-heavy workloads economical. Developers comparing options should note that DeepSeek has since moved to V4-Pro and V4-Flash on its own API, making V3.2 a mature, stable choice for teams that want proven sparse-attention long-context behavior and open-weight flexibility without betting on the newest generation.

NanoGPTdeepseek/deepseek-v3.2deepseek

Open provider page Provider docs

Quick Info

Provider: NanoGPT
Model key: deepseek/deepseek-v3.2
Release date: Dec 1, 2025
Last updated: Dec 1, 2025
Knowledge cutoff: 2024-07
Input modalities
Output modalities
Capabilities

Cost

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

Limits

Input tokens: 163,000 tokens
Output tokens: 65,536 tokens
Context window: 163,000 tokens

Latest news about DeepSeek V3.2

NanoGPT

CoverageBenchmark

Compare GLM-4.6 and DeepSeek-V3.2 across benchmarks, latency, throughput, and real-world performance on DeepInfra to see which open model fits your workloads.

NanoGPT

Coverage

DeepSeek V3.2 is offered through NanoGPT under the model key deepseek/deepseek-v3.2, released December 1, 2025 with a knowledge cutoff of July 2024. The listing on the Sulat models index shows the NanoGPT endpoint exposes text and PDF input with text output, and supports tool calling, structured output, file attachment NanoGPT prices DeepSeek V3.2 at $0.28 per million input tokens and $0.42 per million output tokens, with a 163,000-token input limit, 65,536-token output cap, and a 163,000-token context window. The same DeepSeek V3.2 weights are also served by 27 other providers listed on the page, including OpenRouter, AWS Bedrock, A

NanoGPT

Coverage

Simon Willison's deepseek tag page provides substantive third-party technical characterization of V3.2: DeepSeek released two open-weight MIT-licensed models, DeepSeek-V3.2 and DeepSeek-V3.2-Speciale, both 685B parameters at roughly 690GB, with a linked tech report PDF. V3.2 is positioned as the new flagship distilling The coverage gives independent corroboration of V3.2's scale and open-weight status relevant to NanoGPT developers weighing self-hosting vs API access, and clarifies the capability delta between V3.2 (general flagship) and Speciale (reasoning-only experimental). Given that V4 is now available on DeepSeek's own API and

NanoGPT

CoverageBenchmark

DeepSeek-V3.2-Speciale is a high-compute variant of DeepSeek-V3.2 optimized for maximum reasoning and agentic performance. 131,072 token context window. Includes independent benchmarks from Artificial Analysis.

NanoGPT

Coverage

The DeepSeek API Change Log documents that on 2025-12-01, both deepseek-chat and deepseek-reasoner endpoints were upgraded to DeepSeek-V3.2 — with deepseek-chat mapping to the non-thinking mode and deepseek-reasoner mapping to the thinking mode. A variant named DeepSeek-V3.2-Speciale was also noted as served via a temp The same changelog indicates that on 2026-04-24 DeepSeek introduced V4-Pro and V4-Flash as successors, with the legacy deepseek-chat and deepseek-reasoner model names slated for discontinuation on 2026-07-24. A subsequent V4-Flash API update (2026-07-31) further extended agent benchmarks. As a result, V3.2 is now a leg

NanoGPT

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.

NanoGPT

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