DeepSeek V3.2

Name: DeepSeek V3.2
Price: 0.27 USD
Author: DeepSeek

DeepSeekmid

DeepSeek's updated general-purpose MoE model with 671B total parameters. Offers frontier-competitive quality at ultra-low cost through its efficient MoE architecture. Open-weight and available through the DeepSeek API and numerous third-party providers. The cost-efficiency champion.

Released 2026-02-05Knowledge cutoff: 2025-10

Medium confidence|Updated 74d ago|85% source confidence

Specifications

Context Window

128K tokens

Max Output

16.4K tokens

Input Price

$0.270 / 1M tokens

Output Price

$1.10 / 1M tokens

Latency Tier

Fast (speed score: 7/10)

Capability Profile

Cost Efficiency

10/10

Coding

9/10

Reasoning

8/10

Structured Output

7.5/10

Factuality

7.5/10

Instruction Following

7.5/10

Speed

7/10

Creativity

7/10

Tool Use

7/10

Conversational

7/10

Long Context

6.5/10

Safety & Enterprise

5.5/10

Multimodal

1/10

Feature Support

Vision No

Audio In No

Audio Out No

Video No

Image Generation No

Image Editing No

Function Calling Yes

JSON Mode Yes

Structured Output Yes

Streaming Yes

Reasoning No

Realtime No

Computer Use No

Web Search No

Best Use Cases

Cost-efficient coding and general assistance at 1/10th the cost of frontier models

Batch processing where volume matters and budget is constrained

Self-hosted deployments needing strong quality with open weights

Mathematical reasoning and analytical tasks at budget pricing

Applications where open-weight compliance or customization is needed

Not Ideal For

Multimodal tasks (text only)

Enterprise deployments with strict Western safety/compliance requirements

Tasks requiring real-time information or web grounding

Regulated industries with data sovereignty concerns (data processed in China)

Tasks requiring the highest possible quality regardless of cost

Strengths

Extraordinary cost-to-performance ratio — frontier-competitive at budget pricing

Strong coding abilities that rival models costing 10-20x more

Open weights available for self-hosting and customization

Efficient MoE architecture enables fast inference relative to quality

Good mathematical and analytical reasoning

Weaknesses

Text-only — no multimodal capabilities

Safety alignment reflects Chinese regulatory requirements, not Western enterprise standards

May refuse or deflect questions about Chinese politics, Taiwan, Tiananmen, etc.

128K context is modest compared to the 1M+ models now available

API availability can be inconsistent during high-demand periods

Data may be processed in China — evaluate compliance requirements carefully

Edge Cases & Notes

Chinese content censorship is hard-coded and cannot be removed through prompting

MoE architecture means 671B parameters but much less actual compute per token

Third-party providers (Together, Fireworks) may have different censorship behavior than DeepSeek's own API

V3.2 is a meaningful improvement over V3 on coding and instruction following

Provider Notes

Available through DeepSeek's API and many third-party providers. Self-hosting requires significant infrastructure (multi-GPU). Third-party providers may offer better reliability than DeepSeek's direct API.

Benchmarks

MMLU88.2%

HumanEval90%

Arena Elo1320

Benchmark Notes

MMLU-Pro 88.2%. HumanEval 90%. Competitive with much more expensive models. SWE-bench ~50%. Best cost-adjusted performance of any model. Arena Elo of 1320 is remarkable for the price.

Research Meta

Last Evaluated

2026-03-15

Source Confidence

85%

Evaluation Method

Public benchmarks, cost-quality Pareto analysis, SWE-bench, coding evaluations

Needs Re-evaluation

Sources

DeepSeek V3.2 technical report
LMSYS Chatbot Arena
Open LLM Leaderboard
SWE-bench leaderboard

Continue exploring

Route a prompt

See how DeepSeek V3.2 ranks

Compare models

Side-by-side analysis

Browse registry

Explore all 24 models