CohereLanguage model
Aya Expanse 32B
Aya Expanse 32B has no overall index value yet because there are too few independent benchmark results.
33out of 100-1 vs. last month
based on search, usage and attention
Summary
Aya Expanse 32B is Cohere's open text model for 23 languages. The checkpoint comes from Aya research combining a pretrained Command model with multilingual preference training, safety tuning, and model merging. Its clear focus is cross-language instruction following rather than tool control.
Aya Expanse 32B is the larger family variant, with 32 billion parameters and a 128,000-token context. Aya Expanse 8B is the smaller release. Both weight releases use CC-BY-NC-4.0, which rules out unrestricted commercial use.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 58 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.
Discriminative average (Show measurement, test conditions, and source)
- Source value
- 72.6
- Score
- 72.6
- Metric
- average accuracy over XWinograd, XCOPA and XStoryCloze
- Unit
- %
- Benchmark version
- Aya Expanse paper; 0-shot; 23-language evaluation
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse authors' multilingual evaluation paper
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Expanse 32B results. Global-MMLU is 5-shot; discriminative tasks are zero-shot; MGSM is 5-shot CoT/strict-match; translation uses FLORES-200 and xCOMET-XL.
- Context
- The source is the model authors' paper. It does not disclose every generation detail or evaluation date; no later Aya checkpoint is implied.
Global-MMLU (Show measurement, test conditions, and source)
- Source value
- 66.9
- Score
- 66.9
- Metric
- accuracy; 5-shot
- Unit
- %
- Benchmark version
- Aya Expanse paper; 23-language multilingual MMLU
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse authors' multilingual evaluation paper
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Expanse 32B results. Global-MMLU is 5-shot; discriminative tasks are zero-shot; MGSM is 5-shot CoT/strict-match; translation uses FLORES-200 and xCOMET-XL.
- Context
- The source is the model authors' paper. It does not disclose every generation detail or evaluation date; no later Aya checkpoint is implied.
MGSM (Show measurement, test conditions, and source)
- Source value
- 73.4
- Score
- 73.4
- Metric
- strict-match accuracy; 5-shot chain-of-thought
- Unit
- %
- Benchmark version
- Aya Expanse paper; 7 languages
- Category
- multilingual
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse authors' multilingual evaluation paper
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Aya Expanse 32B results. Global-MMLU is 5-shot; discriminative tasks are zero-shot; MGSM is 5-shot CoT/strict-match; translation uses FLORES-200 and xCOMET-XL.
- Context
- The source is the model authors' paper. It does not disclose every generation detail or evaluation date; no later Aya checkpoint is implied.
INCLUDE accuracy (Show measurement, test conditions, and source)
- Source value
- 51.4
- Score
- 51.4
- Metric
- accuracy
- Unit
- %
- Category
- multilingual-understanding
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. INCLUDE-base, zero-shot native-language chain-of-thought, 44 languages.
m-ArenaHard win rate vs Gemma 2 27B (Show measurement, test conditions, and source)
- Source value
- 51.8
- Score
- 51.8
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard pairwise comparison, averaged across 23 languages; GPT-4o-2024-08-06 judge; instruct-tuned competitors.
m-ArenaHard win rate vs Qwen 2.5 32B (Show measurement, test conditions, and source)
- Source value
- 53.6
- Score
- 53.6
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard pairwise comparison, averaged across 23 languages; GPT-4o-2024-08-06 judge; instruct-tuned competitors.
m-ArenaHard win rate vs Mixtral 8x22B (Show measurement, test conditions, and source)
- Source value
- 76.6
- Score
- 76.6
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard pairwise comparison, averaged across 23 languages; GPT-4o-2024-08-06 judge; instruct-tuned competitors.
m-ArenaHard win rate vs Llama 3.1 70B (Show measurement, test conditions, and source)
- Source value
- 54
- Score
- 54
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. m-ArenaHard pairwise comparison, averaged across 23 languages; GPT-4o-2024-08-06 judge; instruct-tuned competitors.
Dolly win rate vs Mixtral 8x22B (Show measurement, test conditions, and source)
- Source value
- 89.9
- Score
- 89.9
- Metric
- pairwise win rate
- Unit
- %
- Category
- open-ended-generation
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Aya Expanse technical report, arXiv:2412.04261v1
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Dolly held-out evaluation, averaged across 23 languages; GPT-4o-2024-08-06 judge; instruct-tuned competitors.
- Context
- The authors describe Dolly prompts as less difficult and note distribution overlap with Dolly training data.
ALARB verdict prediction / Facts Only / correct-verdict rate (Show measurement, test conditions, and source)
- Source value
- 32.9
- Score
- 32.9
- Metric
- Correct-verdict rate
- Unit
- %
- Benchmark version
- ALARB (ArabicNLP 2025), Table 4
- Category
- legal-reasoning
- Direction
- Higher is better
- Sample
- 1329
- Source type
- independent-evaluator
- Evaluator
- ALARB: An Arabic Legal Argument Reasoning Benchmark (ArabicNLP 2025; DOI 10.18653/v1/2025.arabicnlp-main.32)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. aya-expanse-32B, exact model row in Table 4; verdict prediction with only the case facts, without the court's reasoning; the model is asked to reason and predict a verdict; GPT-4o LLM-as-a-judge classifies alignment with the actual verdict; Table 4 correct-verdict rate for Facts Only.
- Context
- Section 5 states that the evaluation of all tasks used a subset of 1,329 legal cases; this reported evaluation set is applied as N for each Table 4 verdict-prediction task. GPT-4o was used as an LLM judge. No API/checkpoint revision, decoding configuration, or confidence interval is published. Partial and incorrect outcome shares are categorical diagnostics, not separate monotonic quality scores, and are not imported.
ALARB verdict prediction / Facts & Reasoning / correct-verdict rate (Show measurement, test conditions, and source)
- Source value
- 70.6
- Score
- 70.6
- Metric
- Correct-verdict rate
- Unit
- %
- Benchmark version
- ALARB (ArabicNLP 2025), Table 4
- Category
- legal-reasoning
- Direction
- Higher is better
- Sample
- 1329
- Source type
- independent-evaluator
- Evaluator
- ALARB: An Arabic Legal Argument Reasoning Benchmark (ArabicNLP 2025; DOI 10.18653/v1/2025.arabicnlp-main.32)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. aya-expanse-32B, exact model row in Table 4; verdict prediction with case facts plus the court's reasoning; predict the final verdict; GPT-4o LLM-as-a-judge classifies alignment with the actual verdict; Table 4 correct-verdict rate for Facts & Reasoning.
- Context
- Section 5 states that the evaluation of all tasks used a subset of 1,329 legal cases; this reported evaluation set is applied as N for each Table 4 verdict-prediction task. GPT-4o was used as an LLM judge. No API/checkpoint revision, decoding configuration, or confidence interval is published. Partial and incorrect outcome shares are categorical diagnostics, not separate monotonic quality scores, and are not imported.
ALARB verdict prediction / Facts & Regulations / correct-verdict rate (Show measurement, test conditions, and source)
- Source value
- 36.3
- Score
- 36.3
- Metric
- Correct-verdict rate
- Unit
- %
- Benchmark version
- ALARB (ArabicNLP 2025), Table 4
- Category
- legal-reasoning
- Direction
- Higher is better
- Sample
- 1329
- Source type
- independent-evaluator
- Evaluator
- ALARB: An Arabic Legal Argument Reasoning Benchmark (ArabicNLP 2025; DOI 10.18653/v1/2025.arabicnlp-main.32)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. aya-expanse-32B, exact model row in Table 4; verdict prediction with case facts plus relevant legal articles; reason and predict a verdict; GPT-4o LLM-as-a-judge classifies alignment with the actual verdict; Table 4 correct-verdict rate for Facts & Regulations.
- Context
- Section 5 states that the evaluation of all tasks used a subset of 1,329 legal cases; this reported evaluation set is applied as N for each Table 4 verdict-prediction task. GPT-4o was used as an LLM judge. No API/checkpoint revision, decoding configuration, or confidence interval is published. Partial and incorrect outcome shares are categorical diagnostics, not separate monotonic quality scores, and are not imported.
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
Aya Expanse 32BMixtral 8x22B 33Amazon Nova Pro 33MiMo V2.6 Pro UltraSpeed 33Nemotron Mini 4B Instruct 33
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 Aya Expanse 32B, chosen by model line, developer, size, price and release date.
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 Aya Expanse 32B
How large is the context window of Aya Expanse 32B?
128,000 tokens. That is how much text Aya Expanse 32B reads at once, about 190 A4 pages.
Page count as a rule of thumb from Artificial Analysis
Can I run Aya Expanse 32B locally?
Yes, the weights are open.
Stated by Cohere
- Open weights
- yes
- License
- CC BY-NC-4.0
- Parameters
- 32B
- active per token
- n/a
Sources and data date
Show all 22 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
| Evaluator | Test series | Composite indices | Recorded active measurements |
|---|---|---|---|
| LMArena | 0 | 0 | 1 |
Without vendor claims and historical or disputed results; those stay labelled in the single values. 1 measurements are not yet assigned to a test series.
| 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. 14 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | evaluation-details Note The authors' paper specifies datasets, shots and language sets, but not all decoding, checkpoint-revision and run-date details. |
| Not published | serving route and immutable model revision Original wording The result filename identifies the 32B model family/path, but the captured commit does not state serving route or an immutable evaluated checkpoint revision. |
| Not published | evaluation N and confidence interval Original wording No per-result evaluated sample count or confidence interval is published in the captured result rows. acc_norm_stderr is a standard error, not a CI. |
| Unresolved | unit and score direction Original wording The captured current result rows expose acc_norm values but no explicit scale/unit or direction setting; unit remains unknown and higherIsBetter remains null. |
| Not published | evaluation split and inference settings Original wording The captured commit does not establish split, few-shot count, prompt, or decoding settings for these accepted rows. |
| Additional source | Hugging Face (retrieved October 17, 2026) · Cohere Aya Expanse 32B model card · Editorial description reviewed October 2, 2026 |
| Additional source | Aya Expanse authors' multilingual evaluation paper (retrieved October 15, 2026) |
| Additional source | TIGER-Lab official MMLU-Pro leaderboard results.csv (retrieved October 15, 2026) |
| Additional source | Aya Expanse technical report, arXiv:2412.04261v1 (retrieved October 17, 2026) |
| Additional source | Marco-Bench-MIF paper (arXiv:2507.11882v1) (retrieved October 17, 2026) |
| Additional source | ALARB: An Arabic Legal Argument Reasoning Benchmark (ArabicNLP 2025; DOI 10.18653/v1/2025.arabicnlp-main.32) (retrieved October 17, 2026) |
| Additional source | Kids Multilingual Benchmark authors (Batuhan Aktas, Yuvraj, Fatih Buğra Akdoğan), hosted on Cohere Labs Community Blog (retrieved October 17, 2026) |
| Additional source | MIDAS (EMNLP 2025 paper) (retrieved October 18, 2026) |
| Additional source | Polyglot Teachers (arXiv:2604.11290v3) (retrieved October 18, 2026) |
| Additional source | Open Arabic LLM Leaderboard (OALL), v2_results (retrieved October 18, 2026) |
| Additional source | Arena English Text leaderboard (retrieved October 21, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)12 |
| Additional source | Cohere Aya Expanse announcement · 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 |