The AI arena is free today

Open Superagent
DeepSeekReleased on Dec 1, 2025

DeepSeek-V3.2 (Non-thinking): Benchmarks, Pricing & Context Window

DeepSeek-V3.2 (Non-thinking) is a language model from DeepSeek, released in December 2025, with a 131K-token context window, and pricing from $0.280/M input and $0.420/M output.

DeepSeek-V3.2 in non-thinking mode. A powerful language model with 685B parameters using DeepSeek Sparse Attention (DSA) for efficient long-context processing. Supports JSON output, tool calls, and chat prefix completion. This is the

Input
Text
Output
Text
Speed
Latency p95300ms
Throughput P95100char/s
Cost
$0.29/ 1M · 20:1 in:out
$0.28 in · $0.42 out

DeepSeek-V3.2 (Non-thinking) pricing

Providers

DeepSeek-V3.2 (Non-thinking) starts at $0.280 per million input tokens and $0.420 per million output tokens via DeepSeek.

ProviderInput $/MCached input $/MOutput $/MContext in / outTTFT p95 sOutput p5 c/sModalities in / out
DeepSeek logoDeepSeek
$0.280—$0.420131.1K/8.2K
0.30
—
/

Cached input is the discounted price for prompt tokens served from a provider cache. TTFT is time to first token. Output is characters per second; p5 is the sustained floor exceeded by 95% of observed requests.

DeepSeek-V3.2 (Non-thinking) model size

DeepSeek-V3.2 (Non-thinking) has 685 billion parameters. See how it compares to other models in the same parameter range.

Parameters
685BMoE
Frontier (200B+)
685B
1B7B70B405B

DeepSeek-V3.2 (Non-thinking) context window

Input and output token limits for DeepSeek-V3.2 (Non-thinking), plus how it ranks on long-context understanding.

InputOutput
131Ktokens
8Ktokens
≈ 197 pages of text
131K
8K128K1M

Try now

huggle
DeepSeek-V3.2 (Non-thinking)in Huggle

Make it with
DeepSeek-V3.2 (Non-thinking).

DeepSeek-V3.2 (Non-thinking)

DeepSeek-V3.2 (Non-thinking) latency

DeepSeek-V3.2 (Non-thinking) time to first token, sustained output throughput, and failed-request rate from live model usage over the trailing 7 days.

DeepSeek-V3.2 (Non-thinking) examples

Recent arena outputs from DeepSeek-V3.2 (Non-thinking), picked from the highest-ranked matchups.

DeepSeek-V3.2 (Non-thinking) license

DeepSeek-V3.2 (Non-thinking) is released under the MIT license, which permits commercial use, has 685.0B parameters.

License
MIT
Commercial use allowed
Parameters
685.0B

MIT License - allows commercial use

DeepSeek-V3.2 (Non-thinking) resources

Official sources for DeepSeek-V3.2 (Non-thinking): provider documentation, official playground, paper or system card, model weights.

FAQ

Common questions about DeepSeek-V3.2 (Non-thinking).

When was DeepSeek-V3.2 (Non-thinking) released?

DeepSeek-V3.2 (Non-thinking) was released on December 1, 2025 by DeepSeek. This is the official DeepSeek-V3.2 (Non-thinking) release date tracked on LLM Stats.

How much does DeepSeek-V3.2 (Non-thinking) cost?

DeepSeek-V3.2 (Non-thinking) pricing starts at $0.28 per million input tokens and $0.42 per million output tokens via DeepSeek, the lowest price among tracked providers.

How big is DeepSeek-V3.2 (Non-thinking)?

DeepSeek-V3.2 (Non-thinking) has 685 billion parameters. It ships as an open-weight model, so you can download and run it on your own hardware.

Who created DeepSeek-V3.2 (Non-thinking)?

DeepSeek-V3.2 (Non-thinking) was created by DeepSeek.

What is the license for DeepSeek-V3.2 (Non-thinking)?

DeepSeek-V3.2 (Non-thinking) is released under the MIT license. This is an open-source / open-weight license that permits self-hosting.

What is DeepSeek-V3.2 (Non-thinking) latency?

DeepSeek-V3.2 (Non-thinking) p95 time to first token is 0.30 seconds via DeepSeek over the trailing 7 days. Lower time to first token means the model begins responding sooner for chat, agents and model workloads.

Where can I use DeepSeek-V3.2 (Non-thinking)?

DeepSeek-V3.2 (Non-thinking) is available through 1 provider including DeepSeek.

Where is the DeepSeek-V3.2 (Non-thinking) paper or technical report?

DeepSeek-V3.2 (Non-thinking) has a paper or technical report available at https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf. Use that source for architecture, training, release and evaluation details.