Mistral AILanguage model
Pixtral 12B
Pixtral 12B has no overall index value yet because there are too few independent benchmark results.
PopularityRank 526 of 785
31out of 100same as last month
based on search, usage and attention
Summary
What is distinctive about Pixtral 12B is the combination of a 400M vision encoder with a 12B text decoder from the Nemo line. The open multimodal model accepts images of varying size and aspect ratio, multiple images per request, and text with 128,000 tokens of context. The release covers image questions, documents, diagrams, and visual coding. Its weights use Apache 2.0.
Mistral introduced Pixtral 12B as a multimodal replacement for Mistral Nemo; Pixtral Large (latest) is the larger sibling. Mistral’s current model page lists Pixtral 12B as deprecated.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 56 published benchmark results
Page 1 of 5
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.
MMMU (Show measurement, test conditions, and source)
- Source value
- 52.5
- Score
- 52.5
- Metric
- CoT
- Unit
- %
- Category
- vision
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Pixtral 12B model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Pixtral 12B row; CoT is specified. The separate release report states common prompts and a shared evaluation harness; task date and decoding settings are not specified.
MMBench v1.1 dev, Polish (Show measurement, test conditions, and source)
- Source value
- 82.06
- Score
- 82.06
- Metric
- accuracy
- Unit
- %
- Benchmark version
- ACL Anthology 2026.eacl-srw.44
- Category
- multilingual-multimodal
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Annotation-Efficient Vision-Language Model Adaptation to the Polish Language Using LLaVA Framework
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409 on the MMBench v1.1 development set. Accuracy is reported separately for Polish and English. The paper gives no model-specific denominator or confidence interval in the cited table.
- Context
- No model-specific N or CI is published. The MMBench values are not combined with prior generic Pixtral benchmark-card results.
MMBench v1.1 dev, English (Show measurement, test conditions, and source)
- Source value
- 84.31
- Score
- 84.31
- Metric
- accuracy
- Unit
- %
- Benchmark version
- ACL Anthology 2026.eacl-srw.44
- Category
- multilingual-multimodal
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Annotation-Efficient Vision-Language Model Adaptation to the Polish Language Using LLaVA Framework
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409 on the MMBench v1.1 development set. Accuracy is reported separately for Polish and English. The paper gives no model-specific denominator or confidence interval in the cited table.
- Context
- No model-specific N or CI is published. The MMBench values are not combined with prior generic Pixtral benchmark-card results.
AfriMMLU English accuracy: Total (Show measurement, test conditions, and source)
- Source value
- 62
- Score
- 62
- Metric
- zero-shot MCQA accuracy, Total
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; Total category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU English accuracy: Elementary mathematics (Show measurement, test conditions, and source)
- Source value
- 41
- Score
- 41
- Metric
- zero-shot MCQA accuracy, Elementary mathematics
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; Elementary mathematics category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU English accuracy: Global facts (Show measurement, test conditions, and source)
- Source value
- 43
- Score
- 43
- Metric
- zero-shot MCQA accuracy, Global facts
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; Global facts category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU English accuracy: High-school geography (Show measurement, test conditions, and source)
- Source value
- 78
- Score
- 78
- Metric
- zero-shot MCQA accuracy, High-school geography
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; High-school geography category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU English accuracy: High-school macroeconomics (Show measurement, test conditions, and source)
- Source value
- 73
- Score
- 73
- Metric
- zero-shot MCQA accuracy, High-school macroeconomics
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; High-school macroeconomics category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU English accuracy: International law (Show measurement, test conditions, and source)
- Source value
- 76
- Score
- 76
- Metric
- zero-shot MCQA accuracy, International law
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; English-language baseline; International law category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU AfricanLanguagesMean accuracy: Total (Show measurement, test conditions, and source)
- Source value
- 33
- Score
- 33
- Metric
- zero-shot MCQA accuracy, Total
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; mean accuracy across all 16 African languages; Total category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU AfricanLanguagesMean accuracy: Elementary mathematics (Show measurement, test conditions, and source)
- Source value
- 35
- Score
- 35
- Metric
- zero-shot MCQA accuracy, Elementary mathematics
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; mean accuracy across all 16 African languages; Elementary mathematics category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
AfriMMLU AfricanLanguagesMean accuracy: Global facts (Show measurement, test conditions, and source)
- Source value
- 31
- Score
- 31
- Metric
- zero-shot MCQA accuracy, Global facts
- Unit
- %
- Benchmark version
- AfriMMLU; ACL AfricaNLP 2026 (2026.africanlp-main.10)
- Category
- multilingual-multiple-choice-question-answering
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- The Token Tax: Systematic Bias in Multilingual Tokenization
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Pixtral-12B-2409; zero-shot multiple-choice question answering on AfriMMLU; mean accuracy across all 16 African languages; Global facts category; score read as a rounded integer from Figure 1.
- Context
- Figure 1 reports rounded integer cells. Per-cell question denominators and confidence intervals are not published, and the aggregation rule for Total is not specified. The African-language row is the mean across 16 languages, not a pooled per-item rate. These AfriMMLU results do not establish performance for every African language or task. The figure does not state an immutable inference/API route or checkpoint digest.
Page 1 of 5
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
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 7, 2026
Pixtral 12BSarvam 105B 31Agnes 3.0 Flash Preview 30Qwen 3 VL 8B Thinking 30Yi-1.5 34B 30
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Pixtral 12B, chosen by model line, developer, size, price and release date.
- Mistral 7B v0.3Mistral AI · same developer
- Ministral 3 14BMistral AI · same developer
- Codestral-22B-v0.1Mistral AI · same developer
- Codestral (latest)Mistral AI · same developer
- Magistral SmallMistral AI · same developer
- Devstral SmallMistral AI · same developer
- Mixtral 8x7BMistral AI · same developer
- Mistral Large 2407Mistral AI · same developer
All specifications
Architecture
Input and output
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
Frequently asked questions about Pixtral 12B
What is the knowledge cutoff of Pixtral 12B?
September 2024, according to Mistral AI. That is the month of release.
Stated by Mistral AI
SepKnowledge cutoffSeptember 2024ReleasedSep 11, 2024
How large is the context window of Pixtral 12B?
128,000 tokens. That is how much text Pixtral 12B reads at once, about 190 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
GLM-4.5131K
GPT OSS 120B131K
Pixtral 12BRank 98 of 103128K
DeepSeek-V3.2128K
DeepSeek-R1128K
Can I run Pixtral 12B locally?
Yes, the weights are open.
Stated by Mistral AI
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 12B
- active per token
- n/a
Sources and data date
Show all 19 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 |
|---|---|
| Input | Documented modalities. Retrieved October 27, 2026. |
| Output | Documented modalities. Retrieved October 27, 2026. |
| Research status | Research date October 15, 2026. 11 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | score-units-and-run-details Note The card names metrics and selected prompting modes; most raw values have no unit label, and task dates, samples, and per-task prompts are absent. |
| Not published | if-eval-scale-variant-and-sample-count Note The current card publishes Text IF-Eval 61.3 and MM IF-Eval 52.7 but does not name the exact IF-Eval variant, unit scale, prompts, sample count, or confidence intervals. |
| Unresolved | card-paper-score-conflicts Note The current model card reports MMMU 52.5 and MathVista 58.0, while the Pixtral paper reports MMMU 52.0 and MathVista 58.3. These existing rows were not replaced or reconciled. |
| Not published | model-specific-denominator Original wording The cited MMBench table does not state a model-specific denominator or confidence interval. |
| Unresolved | overall-accuracy Original wording The abstract says Pixtral 12B scored 51.7%, while Table II prints 0.5 as Total and Table III reports 60 correct of 120 questions. The overall value is excluded; the eight included task rows are individually supported by Table III counts. |
| Additional source | Hugging Face (retrieved October 15, 2026) · Pixtral 12B model card · Editorial description reviewed October 2, 2026 |
| Additional source | Mistral Pixtral 12B model card (retrieved October 17, 2026) |
| Additional source | Pixtral 12B paper reporting the LMSys Vision Arena snapshot (retrieved October 17, 2026) |
| Additional source | Annotation-Efficient Vision-Language Model Adaptation to the Polish Language Using LLaVA Framework (retrieved October 17, 2026) |
| Additional source | The Token Tax: Systematic Bias in Multilingual Tokenization (retrieved October 17, 2026) |
| Additional source | Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases (retrieved October 18, 2026) |
| Additional source | OCRBench v2 paper, NeurIPS Datasets and Benchmarks Track (retrieved October 18, 2026) |
| Additional source | Tonmoy et al., Vision-Language Models for Criterion-Level Grading of Handwritten Examinations in Outcome-Based Education (retrieved October 18, 2026) |
| Additional source | Authors of Visual Reasoning Evaluation of Grok, Deepseek’s Janus, Gemini, Qwen, Mistral, and ChatGPT (arXiv:2502.16428) (retrieved October 19, 2026) |
| Additional source | Bechard et al., StarFlow: Generating Structured Workflow Outputs From Sketch Images (arXiv:2503.21889v2, Table 2) (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026) · Pixtral 12B announcement · Editorial description reviewed October 2, 2026 |
| Additional source | Mistral model overview · Editorial description reviewed October 2, 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 |