The AI arena is free today

Open Superagent

DeepSeek-V3.2 (Thinking) vs MAI-Thinking-1

DeepSeek-V3.2 (Thinking) and MAI-Thinking-1 are closely matched at 32.7 and 33.1 on the LLM Stats Score.

DeepSeek · Microsoft · Updated for 2026

Which is better?

DeepSeek-V3.2 (Thinking) and MAI-Thinking-1 are closely matched on the overall LLM Stats Score at 32.7 and 33.1.

In the 5 individual benchmarks reported for both models, MAI-Thinking-1 wins 4; this is a narrower head-to-head signal than the composite indexes.

Based on current LLM Stats indexes, shared benchmarks, pricing, and model metadata for 2026.

Choose DeepSeek-V3.2 (Thinking)

  • you need open weights you can self-host or fine-tune

Choose MAI-Thinking-1

  • you value its reported benchmark strengths — it wins 4 of 5 exact shared results
  • you want the most recent training data — it shipped Jun 2026

At a glance

The differences that matter most.

DeepSeek-V3.2 (Thinking)
DeepSeek
MAI-Thinking-1
Microsoft
Core performance indexes
LLM Stats Score
32.7
#102
33.1
#100
Reasoning
32.7
#100
33.9
#91
Coding
22.9
#77
19.4
#103
Agents
11.2
#109
12.8
#96
Cost, coverage & limits
Benchmark wins
1 of 5
4 of 5
Input price
$0.28 / M
— / M
Output price
$0.42 / M
— / M
Context window
131,072
—

Capability indexes

Additional strengths measured across groups of related public benchmarks

2 shared
Index
DeepSeek-V3.2 (Thinking)
MAI-Thinking-1
Math
30.3#76
33.7#52
Tool use
10.0#128
10.7#118
Conservative TrueSkill rating · higher is betterHow scores work

Individual benchmarks

14 reported for DeepSeek-V3.2 (Thinking) · 23 for MAI-Thinking-1

5 shared

DeepSeek-V3.2 (Thinking) outperforms in 1 benchmarks (Terminal-Bench 2.0), while MAI-Thinking-1 is better at 3 benchmarks (AIME 2025, GPQA, SWE-Bench Verified).

MAI-Thinking-1 has a slight edge in benchmark performance.

Wed Sep 09 2026 • llm-stats.com

Human preference

Blind head-to-head votes and playground preference scores

Model Size

Parameter count comparison

315.0B diff

MAI-Thinking-1 has 315.0B more parameters than DeepSeek-V3.2 (Thinking), making it 46.0% larger.

DeepSeek
DeepSeek-V3.2 (Thinking)
685.0Bparameters
Microsoft
MAI-Thinking-1
1.0Tparameters
685.0B
DeepSeek-V3.2 (Thinking)
1000.0B
MAI-Thinking-1

Context Window

Maximum input and output token capacity

Only DeepSeek-V3.2 (Thinking) specifies input context (131,072 tokens). Only DeepSeek-V3.2 (Thinking) specifies output context (65,536 tokens).

DeepSeek
DeepSeek-V3.2 (Thinking)
Input131,072 tokens
Output65,536 tokens
Microsoft
MAI-Thinking-1
Input- tokens
Output- tokens
Wed Sep 09 2026 • llm-stats.com

License

Usage and distribution terms

DeepSeek-V3.2 (Thinking) is licensed under MIT, while MAI-Thinking-1 uses a proprietary license.

License differences may affect how you can use these models in commercial or open-source projects.

DeepSeek-V3.2 (Thinking)

MIT

Open weights

MAI-Thinking-1

Proprietary

Closed source

Release Timeline

When each model was launched

DeepSeek-V3.2 (Thinking) was released on 2025-12-01, while MAI-Thinking-1 was released on 2026-06-02.

MAI-Thinking-1 is 6 months newer than DeepSeek-V3.2 (Thinking).

DeepSeek-V3.2 (Thinking)

Dec 1, 2025

9 months ago

MAI-Thinking-1

Jun 2, 2026

3 months ago

6mo newer

Knowledge Cutoff

When training data ends

Neither model specifies a knowledge cutoff date.

Unable to compare the recency of their training data.

No cutoff dates available

Outputs Comparison

Notice missing or incorrect data?Start an Issue discussion→

Judge for yourself.

Run your own prompts against DeepSeek-V3.2 (Thinking) and MAI-Thinking-1 side-by-side, then vote on the output you prefer.

DeepSeek-V3.2 (Thinking)
✓ Preferred
MAI-Thinking-1
Open in Playground

FAQ

Common questions about DeepSeek-V3.2 (Thinking) vs MAI-Thinking-1.

Which is better, DeepSeek-V3.2 (Thinking) or MAI-Thinking-1?

DeepSeek-V3.2 (Thinking) and MAI-Thinking-1 are closely matched on the LLM Stats Score at 32.7 and 33.1. DeepSeek-V3.2 (Thinking) is made by DeepSeek and MAI-Thinking-1 is made by Microsoft. The best choice depends on your use case — compare their capability indexes, individual benchmarks, pricing, and limits above.

How does DeepSeek-V3.2 (Thinking) compare to MAI-Thinking-1 in benchmarks?

DeepSeek-V3.2 (Thinking) scores AIME 2025: 93.1%, HMMT 2025: 90.2%, MMLU-Pro: 85.0%, LiveCodeBench: 83.3%, GPQA: 82.4%. MAI-Thinking-1 scores LongFact: 98.0%, AIME 2025: 97.0%, AIME 2026: 94.5%, GraphWalks: 90.0%, AIR-Bench: 88.0%.

What are the context window sizes for DeepSeek-V3.2 (Thinking) and MAI-Thinking-1?

DeepSeek-V3.2 (Thinking) supports 131K tokens and MAI-Thinking-1 supports an unknown number of tokens. A larger context window lets you process longer documents, conversations, or codebases in a single request.

What are the main differences between DeepSeek-V3.2 (Thinking) and MAI-Thinking-1?

Key differences include LLM Stats Score (32.7 vs 33.1), licensing (MIT vs Proprietary). See the full comparison above for benchmark-by-benchmark results.

Who makes DeepSeek-V3.2 (Thinking) and MAI-Thinking-1?

DeepSeek-V3.2 (Thinking) is developed by DeepSeek and MAI-Thinking-1 is developed by Microsoft.

Related comparisons