Microsoft AILanguage modelActive
MAI-Code-1.1-Flash
MAI-Code-1.1-Flash has no overall index value yet because there are too few independent benchmark results.
PopularityRank 648 of 875
1out of 100same as last month
Summary
MAI-Code-1.1-Flash reduces token use on coding tasks. Microsoft reports 25% fewer tokens per task and 25% faster token streaming in GitHub Copilot compared with version 1.0. Its post-training focuses in part on CLI tasks and .NET.
Microsoft put the version into production in GitHub Copilot on August 11, 2026. The release offers neither open weights nor a separate public API endpoint.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 12 published benchmark results
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.
SWE-bench Verified pass rate (Show measurement, test conditions, and source)
- Source value
- 72.6
- Score
- 72.6
- Metric
- SWE-bench Verified pass rate
- Unit
- %
- Category
- coding
- Direction
- Higher is better
- Sample
- 500
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. GitHub Copilot production harness; VS Code repository context, tool calls, and verification; model-card page 5.
- Context
- Vendor-reported, harness-specific result. Reasoning effort, sample count, and evaluation date are unspecified; not a shared comparison cohort.
Terminal-Bench 2.1 pass rate (Show measurement, test conditions, and source)
- Source value
- 62.9
- Score
- 62.9
- Metric
- Terminal-Bench 2.1 pass rate
- Unit
- %
- Category
- coding
- Direction
- Higher is better
- Sample
- 89
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. GitHub Copilot production harness; VS Code repository context, tool calls, and verification; model-card page 5.
- Context
- Vendor-reported, harness-specific result. Reasoning effort, sample count, and evaluation date are unspecified; not a shared comparison cohort.
Text2WebApp pass rate (Show measurement, test conditions, and source)
- Source value
- 74.1
- Score
- 74.1
- Metric
- Text2WebApp pass rate
- Unit
- %
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Internal visual-coding evaluation in the GitHub Copilot production harness; text specification to web application; model-card page 5.
- Context
- Vendor-reported internal benchmark. Public tasks, grading protocol, sample count, reasoning effort, and run date are unspecified.
ScreenShot2WebApp pass rate (Show measurement, test conditions, and source)
- Source value
- 42.1
- Score
- 42.1
- Metric
- ScreenShot2WebApp pass rate
- Unit
- %
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Internal visual-coding evaluation in the GitHub Copilot production harness; screenshot to web application; model-card page 5.
- Context
- Vendor-reported internal benchmark. Public tasks, grading protocol, sample count, reasoning effort, and run date are unspecified.
Vision2Web Level 3 pass rate (Show measurement, test conditions, and source)
- Source value
- 11.5
- Score
- 11.5
- Metric
- Vision2Web Level 3 pass rate
- Unit
- %
- Category
- coding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Source-specific Vision2Web Level 3 evaluation in the GitHub Copilot production harness; model-card page 5.
- Context
- Vendor-reported visual-coding benchmark. This card does not provide public tasks, grading protocol, sample count, reasoning effort, or run date, nor explicitly label Vision2Web internal.
Terminal-Bench 2.1 (GitHub Copilot CLI), reported improvement (Show measurement, test conditions, and source)
- Source value
- 22
- Score
- 22
- Metric
- Source-stated improvement magnitude; relative percent versus percentage-point basis unspecified
- Unit
- %
- Benchmark version
- Terminal-Bench 2.1; GitHub Copilot CLI; Microsoft AI article dated 2026-10-07
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash: Better, faster, at a quarter of the cost (Microsoft AI, 2026-10-07)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact reported model MAI-Code-1.1-Flash; source says the 22% improvement was measured on Terminal-Bench 2.1 in GitHub Copilot CLI. The article says it is relative to the model launched in June 2026, but does not identify that comparator by exact model ID. It does not publish an absolute pass rate, benchmark run count, task denominator, API route, immutable serving revision, or reasoning-effort setting.
- Context
- The source says ‘22% improvement’ but does not define whether that is a relative percent change or a percentage-point change. It does not name the June comparator or publish the underlying absolute scores, sample count, or confidence interval. This is retained as a source-stated improvement, not converted into an absolute benchmark score.
SWE-Bench Verified (Show measurement, test conditions, and source)
- Source value
- 70.8
- Score
- 70.8
- Metric
- Pass rate
- Unit
- %
- Direction
- Higher is better
- Sample
- 500
- Source type
- vendor-reported
- Evaluator
- Bringing local models and sandboxed tools to Windows and GitHub Copilot
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Vendor-reported result. MAI Code 1.1 Flash Quantized on Device, as labeled in Microsoft's table. The article reports 53 GB and 80% smaller than the BF16 cloud variant for the on-device version, but the table does not state exact quantization precision; no immutable checkpoint digest or dataset split revision is given.
- Context
- This is Microsoft's on-device quantized model column. The article reports 53 GB and 80% smaller than the BF16 cloud variant, but gives no exact quantization precision, immutable checkpoint digest, dataset split revision, or row-specific run date.
Terminal-Bench 2.1 (Show measurement, test conditions, and source)
- Source value
- 66.29
- Score
- 66.29
- Metric
- Pass rate
- Unit
- %
- Direction
- Higher is better
- Sample
- 89
- Source type
- vendor-reported
- Evaluator
- Bringing local models and sandboxed tools to Windows and GitHub Copilot
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Vendor-reported result. MAI Code 1.1 Flash Quantized on Device, as labeled in Microsoft's table. The article reports 53 GB and 80% smaller than the BF16 cloud variant for the on-device version, but the table does not state exact quantization precision; no immutable checkpoint digest or dataset split revision is given.
- Context
- This is Microsoft's on-device quantized model column. The article reports 53 GB and 80% smaller than the BF16 cloud variant, but gives no exact quantization precision, immutable checkpoint digest, dataset split revision, or row-specific run date.
Prompt-processing throughput (Show measurement, test conditions, and source)
- Source value
- 923.5
- Score
- 923.5
- Metric
- Prompt-processing throughput at 64k context
- Unit
- tokens/s
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bringing local models and sandboxed tools to Windows and GitHub Copilot
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Vendor-reported result. Microsoft's first-shipping local MAI Code 1.1 Flash version on Surface Laptop Ultra; prompt-processing throughput at 64k context. The L44 sentence does not state exact quantization precision, runtime version, batch size, or timed input length. The footnote for the separately captioned decode-throughput chart specifies a test on October 5, 2026 with mixed-precision quantization at approximately 3.3 bits per weight, DFlash2 sliding-window speculative decoding, and a Windows ARM64 llama.cpp CUDA runtime; those results are for a synthetic code-generation workload. The source does not bind that decode setup to these prompt-processing values.
- Context
- Device-specific vendor result. Microsoft publishes the decode-throughput chart and its setup separately; the footnote labels those as decode results and does not bind that setup to the 64k prompt-processing value. No checkpoint digest is given, so this is not comparable to provider endpoint throughput.
Prompt-processing throughput (Show measurement, test conditions, and source)
- Source value
- 769.8
- Score
- 769.8
- Metric
- Prompt-processing throughput at 128k context
- Unit
- tokens/s
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Bringing local models and sandboxed tools to Windows and GitHub Copilot
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Vendor-reported result. Microsoft's first-shipping local MAI Code 1.1 Flash version on Surface Laptop Ultra; prompt-processing throughput at 128k context. The L44 sentence does not state exact quantization precision, runtime version, batch size, or timed input length. The footnote for the separately captioned decode-throughput chart specifies a test on October 5, 2026 with mixed-precision quantization at approximately 3.3 bits per weight, DFlash2 sliding-window speculative decoding, and a Windows ARM64 llama.cpp CUDA runtime; those results are for a synthetic code-generation workload. The source does not bind that decode setup to these prompt-processing values.
- Context
- Device-specific vendor result. Microsoft publishes the decode-throughput chart and its setup separately; the footnote labels those as decode results and does not bind that setup to the 128k prompt-processing value. No checkpoint digest is given, so this is not comparable to provider endpoint throughput.
SWE-Bench Verified, average tokens per completed task (Show measurement, test conditions, and source)
- Source value
- 8600
- Score
- 8600
- Metric
- Average tokens used per completed task (solution length)
- Unit
- mean-count
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Vendor-reported result. Microsoft reports MAI-Code-1.1-Flash in its model-card table. SWE-Bench Verified was evaluated in the same VS Code-based harness used for production GitHub Copilot workflows, end to end with repository context, tool calls, and verification. The table reports average token use per completed task, with the same harness and settings for compared models.
- Context
- The card defines this as solution length, the average number of tokens used per completed task, and says shorter responses are preferable only when pass rates are comparable. It does not specify which token categories its accounting includes, so this must not be read as a count of model output/completion tokens. The source-rounded 8.6K is normalized here to 8,600, not represented as an unrounded count. The task denominator, confidence interval, and row-specific run date are not reported.
Terminal Bench 2.1, average tokens per completed task (Show measurement, test conditions, and source)
- Source value
- 17000
- Score
- 17000
- Metric
- Average tokens used per completed task (solution length)
- Unit
- mean-count
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- MAI-Code-1.1-Flash Model Card
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Vendor-reported result. Microsoft reports MAI-Code-1.1-Flash in its model-card table. Terminal Bench 2.1 was evaluated in the same VS Code-based harness used for production GitHub Copilot workflows, end to end with repository context, tool calls, and verification. The table reports average token use per completed task, with the same harness and settings for compared models.
- Context
- The card defines this as solution length, the average number of tokens used per completed task, and says shorter responses are preferable only when pass rates are comparable. It does not specify which token categories its accounting includes, so this must not be read as a count of model output/completion tokens. The source-rounded 17.0K is normalized here to 17,000, not represented as an unrounded count. The task denominator, confidence interval, and row-specific run date are not reported.
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
Gemini 4 Argon, Google DeepMind: 92
GPT-6 Astra, OpenAI: 78
Claude Haiku 5.5, Anthropic: 77
Qwen3.5 4B, Alibaba: 76
GPT-5, OpenAI: 73
GPT-6 Sol, OpenAI: 72
Qwen3.5 0.8B, Alibaba: 72
Qwen3.5 2B, Alibaba: 71
Qwen 3.8 27B, Alibaba: 70
Claude Opus 5.5, Anthropic: 70
DeepSeek V4.1 Flash 1017, DeepSeek: 68
Qwen 3.8 Max 0902, Alibaba: 68
Kimi K3, Moonshot AI: 68
DeepSeek-V4-Flash, DeepSeek: 65
GPT-5.6 Terra, OpenAI: 64
Claude Sonnet 5, Anthropic: 63
DiffusionGemma 26B A4B IT, Google: 63
Claude Opus 5, Anthropic: 62
Qwen 3.8 Flash, Alibaba: 62
MAI-Code-1.1-Flash, Microsoft AI: 1
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 5, 2026
MAI-Code-1.1-FlashQwen3.8-Omni-FlashSWE-2Step 5 PreviewMAI-Thinking-1
All specifications
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
Further details
Prices
No Microsoft AI subscription includes this model. Through other providers from ≈ €10.60 a month (GitHub Copilot Pro, with limited credits). 5 plans in Germany.
| Plan | Monthly price | Note |
|---|---|---|
| GitHub Copilot Pro | ≈ €10.60 | with limited credit |
| GitHub Copilot Business | ≈ €20.14 | with limited credit, for teams, per seat |
| GitHub Copilot Enterprise | ≈ €41.35 | with limited credit, for enterprises, per seat |
| GitHub Copilot Pro+ | ≈ €41.35 | with limited credit |
| GitHub Copilot Max | ≈ €106.02 | with limited credit |
Frequently asked questions about MAI-Code-1.1-Flash
Can I run MAI-Code-1.1-Flash locally?
No. Microsoft AI does not publish the weights.
Stated by Microsoft AI
- Open weights
- no
- License
- n/a
- Parameters
- n/a
- active per token
- n/a
Where can I use MAI-Code-1.1-Flash?
Not in a Microsoft AI subscription. Including other providers that makes 5 plans in Germany.
Subscription comparison, prices in Germany incl. VAT
At Microsoft AI
No Microsoft AI subscription includes this model.
Through other providers
5 plans from 1 provider, from ≈ €10.60 a month (GitHub Copilot Pro), with limited credits.
Sources and data date
Show all 9 sources
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 | vision-benchmark-protocol Note The card reports exact pass rates for the three visual-coding tests without public tasks, grading protocol, or sample count. Only Text2WebApp and ScreenShot2WebApp are explicitly labelled internal; that label is not inferred for Vision2Web. |
| Not published | improvement-basis-and-comparator Original wording The article does not define whether 22% denotes relative change or percentage points and does not identify the June comparator by exact model ID. |
| Not published | absolute-score-sample-count-and-reasoning-setup Original wording No absolute Terminal-Bench 2.1 pass rate, task denominator, run count, serving revision, API route, or reasoning effort is reported. |
| Not published | confidence-intervals Original wording The article reports no confidence interval for the improvement. |
| Additional source | Microsoft AI MAI-Code-1.1-Flash announcement (retrieved October 3, 2026; October 4, 2026; October 17, 2026) · Microsoft AI MAI-Code-1.1-Flash announcement · Editorial description reviewed October 4, 2026 |
| Additional source | MAI-Code-1.1-Flash Model Card (retrieved October 4, 2026; October 23, 2026) |
| Additional source | Bringing local models and sandboxed tools to Windows and GitHub Copilot (retrieved October 20, 2026) |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |