Microsoft AILanguage modelPreview
MAI-Thinking-1
MAI-Thinking-1 has no overall index value yet because there are too few independent benchmark results. From the vendor-reported results we estimate about 73.2 points (95% range 53.2 to 93.1). This is not an independent measurement.
n/ano search or usage data
Summary
MAI-Thinking-1 was trained without distillation from other providers’ models, according to Microsoft. The company uses its own traceable training data and executable coding environments where real tests grade outcomes. The MoE model activates 35 billion of roughly 1 trillion parameters.
It has been available in public preview on Microsoft Foundry since August 12, 2026. Microsoft specifies a 256K-token context window.
On this page
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 44 published benchmark results
Page 1 of 4
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
AIME 2025 (Show measurement, test conditions, and source)
- Source value
- 97
- Metric
- AIME 2025
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
AIME 2026 (Show measurement, test conditions, and source)
- Source value
- 94.5
- Metric
- AIME 2026
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
HMMT Feb 2026 (Show measurement, test conditions, and source)
- Source value
- 84.9
- Metric
- HMMT Feb 2026
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 84.2
- Metric
- GPQA Diamond
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
LCB v6 (Show measurement, test conditions, and source)
- Source value
- 87.7
- Metric
- LCB v6
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
Terminal Bench 2.0 (Show measurement, test conditions, and source)
- Source value
- 46
- Metric
- Terminal Bench 2.0
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
SWE-Bench Verified (Show measurement, test conditions, and source)
- Source value
- 73.5
- Metric
- SWE-Bench Verified
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
SWE-Bench Pro (Show measurement, test conditions, and source)
- Source value
- 52.8
- Metric
- SWE-Bench Pro
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Microsoft announcement Table 1 pixels
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Post-trained model evaluation; percentages. Agentic coding uses 256K total context; others max output 256K. Separate blind human preference study has 1,276 tasks.
- Context
- Source-specific observation; it is not a shared comparison cohort. Microsoft self-reported table; peer numbers taken from respective official cards
Human side-by-side preference vs Claude Sonnet 4.6 (win share) (Show measurement, test conditions, and source)
- Source value
- 49
- Metric
- Share of pairwise comparisons won by MAI-Thinking-1
- Unit
- %
- Benchmark version
- Microsoft technical report Table 14; task pool 1,276; comparator as named
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Thinking-1: Building a Hill-Climbing Machine (Microsoft AI technical report)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact reported model name MAI-Thinking-1; vendor human side-by-side evaluation against the named Claude comparator. The task pool comprises 1,276 English tasks, 30% multi-turn, rated by human evaluators; overall preference uses a seven-point scale from -1.5 to +1.5. Model serving revision, API route, reasoning effort, generation settings, and the exact number of paired ratings are not published for this evaluation.
- Context
- The 1,276 count is the published task-pool size, not a confirmed denominator of paired ratings, so sampleCount is omitted. The source prints +/- 0.06 with the mean preference rows but does not define that notation in the reported section; it is not entered as a confidence interval. The source's 6% and 5% tie shares are documented in the capture and note but not imported because the required directional boolean cannot represent a neutral tie rate without implying a quality direction.
Human side-by-side preference vs Claude Sonnet 4.6 (loss share) (Show measurement, test conditions, and source)
- Source value
- 45
- Metric
- Share of pairwise comparisons lost by MAI-Thinking-1
- Unit
- %
- Benchmark version
- Microsoft technical report Table 14; task pool 1,276; comparator as named
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- MAI-Thinking-1: Building a Hill-Climbing Machine (Microsoft AI technical report)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact reported model name MAI-Thinking-1; vendor human side-by-side evaluation against the named Claude comparator. The task pool comprises 1,276 English tasks, 30% multi-turn, rated by human evaluators; overall preference uses a seven-point scale from -1.5 to +1.5. Model serving revision, API route, reasoning effort, generation settings, and the exact number of paired ratings are not published for this evaluation.
- Context
- The 1,276 count is the published task-pool size, not a confirmed denominator of paired ratings, so sampleCount is omitted. The source prints +/- 0.06 with the mean preference rows but does not define that notation in the reported section; it is not entered as a confidence interval. The source's 6% and 5% tie shares are documented in the capture and note but not imported because the required directional boolean cannot represent a neutral tie rate without implying a quality direction.
Human side-by-side preference vs Claude Opus 4.6 (win share) (Show measurement, test conditions, and source)
- Source value
- 43
- Metric
- Share of pairwise comparisons won by MAI-Thinking-1
- Unit
- %
- Benchmark version
- Microsoft technical report Table 14; task pool 1,276; comparator as named
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Thinking-1: Building a Hill-Climbing Machine (Microsoft AI technical report)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact reported model name MAI-Thinking-1; vendor human side-by-side evaluation against the named Claude comparator. The task pool comprises 1,276 English tasks, 30% multi-turn, rated by human evaluators; overall preference uses a seven-point scale from -1.5 to +1.5. Model serving revision, API route, reasoning effort, generation settings, and the exact number of paired ratings are not published for this evaluation.
- Context
- The 1,276 count is the published task-pool size, not a confirmed denominator of paired ratings, so sampleCount is omitted. The source prints +/- 0.06 with the mean preference rows but does not define that notation in the reported section; it is not entered as a confidence interval. The source's 6% and 5% tie shares are documented in the capture and note but not imported because the required directional boolean cannot represent a neutral tie rate without implying a quality direction.
Human side-by-side preference vs Claude Opus 4.6 (loss share) (Show measurement, test conditions, and source)
- Source value
- 52
- Metric
- Share of pairwise comparisons lost by MAI-Thinking-1
- Unit
- %
- Benchmark version
- Microsoft technical report Table 14; task pool 1,276; comparator as named
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- MAI-Thinking-1: Building a Hill-Climbing Machine (Microsoft AI technical report)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact reported model name MAI-Thinking-1; vendor human side-by-side evaluation against the named Claude comparator. The task pool comprises 1,276 English tasks, 30% multi-turn, rated by human evaluators; overall preference uses a seven-point scale from -1.5 to +1.5. Model serving revision, API route, reasoning effort, generation settings, and the exact number of paired ratings are not published for this evaluation.
- Context
- The 1,276 count is the published task-pool size, not a confirmed denominator of paired ratings, so sampleCount is omitted. The source prints +/- 0.06 with the mean preference rows but does not define that notation in the reported section; it is not entered as a confidence interval. The source's 6% and 5% tie shares are documented in the capture and note but not imported because the required directional boolean cannot represent a neutral tie rate without implying a quality direction.
Page 1 of 4
Similar models
More models of the same kind as MAI-Thinking-1, chosen by model line, developer, size, price and release date.
All specifications
Architecture
- Parameters
- About 1 trillion total according to Microsoft, 35 billion active (mixture of experts)
- Source: Parameters
Input and output
- Capabilities
- Function calling, developer instructions, and a Chat Completions-compatible API
- Source: Capabilities
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- Shutdown
- None announced
- no source
Further details
Frequently asked questions about MAI-Thinking-1
How good is MAI-Thinking-1 in benchmarks?
Independent results are not yet enough for the overall index. An estimate from 5 vendor-reported results (AIME 2025 97%, GPQA Diamond 84.2%, SWE-bench Pro (public) 52.8%, SWE-bench Verified 73.5%, Terminal-Bench 2.0 46%) gives about 73.2 points, with a 95% range of 53.2 to 93.1. The estimate is calibrated on models with both vendor-reported and independent results. Vendor numbers run clearly above independent measurements on some tests; the calibration offsets that on average.
Vendor estimate, not an overall index value
How large is the context window of MAI-Thinking-1?
256,000 tokens. Separate limits may apply to input and output.
Can I run MAI-Thinking-1 locally?
No. Microsoft AI does not publish the weights.
Stated by Microsoft AI
- Open weights
- no
- License
- n/a
- Parameters
- 1T
- active per token
- 35B
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 4, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | score-scale-and-scored-denominator Note Tables 12 and 19 omit explicit units, per-task scored denominators, and confidence intervals. Table 19 states a four-run average but not run-to-run dispersion. |
| Not published | paired-rating-denominator Original wording The 1,276 number is the task-pool size; the exact number of paired ratings contributing to each outcome/mean is not explicitly reported. |
| Unresolved | plus-minus-notation Original wording Table 14 prints +/- 0.06 beside each mean preference but the cited section does not define whether it is a standard error, interval, or another quantity; it is not encoded as a confidence interval. |
| Not published | serving-revision-and-reasoning-effort Original wording The human evaluation does not publish the MAI-Thinking-1 serving revision, route, reasoning effort, or generation settings. |
| Unresolved | neutral-tie-rate-projection Original wording The paper reports 6% ties versus Sonnet and 5% versus Opus. Those neutral shares are retained in the note/capture but withheld from the schema because higherIsBetter is mandatory and cannot express a neutral direction. |
| Additional source | Microsoft AI MAI-Thinking-1 announcement (retrieved October 4, 2026; October 25, 2026) · Microsoft AI MAI-Thinking-1 announcement · Editorial description reviewed October 4, 2026 |
| Additional source | Microsoft MAI-Thinking-1 technical report (retrieved October 17, 2026) |
| Additional source | MathBench.ai (retrieved October 28, 2026) |
| Additional source | Model metadata (retrieved October 28, 2026) |