MGSM Bengali (Kaggle v1) (MGSM Bengali (Kaggle))
We show this table for reference; we do not rank on it.
Tests mathematical reasoning on the Bengali translation of 250 grade-school math problems from GSM8K.
Accuracy on MGSM Bengali (Kaggle) — 2026-10-20 source capture
We mirror the published accuracy view for MGSM Bengali (Kaggle). Claude Opus 4.8 has the highest reported value at 95.6%. These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. We do not use these results to rank models overall.
Claude Opus 4.8
Evaluation date not reported
Claude Opus 4
Evaluated 2025-07-01
GPT-5
Evaluated 2025-08-19
46 modelsMathematicsStaleDisplay onlyUpdated 2026-10-20 source capture
Accuracy table (46 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 4.8Creator not reportedEvaluation date not reported | 95.6% | Not reported | Unknown |
| 2 | Claude Opus 4Creator not reportedEvaluated 2025-07-01 | 94.4% | Not reported | Unknown |
| 3 | GPT-5Creator not reportedEvaluated 2025-08-19 | 94.0% | Not reported | Unknown |
| 4 | Gemini 2.5 Pro PreviewCreator not reportedEvaluated 2025-06-30 | 92.8% | Not reported | Unknown |
| 5 | Claude Opus 4.1Creator not reportedEvaluated 2025-08-20 | 92.8% | Not reported | Unknown |
| 6 | Gemma 4 26B A4BCreator not reportedEvaluation date not reported | 92.4% | Not reported | Unknown |
| 7 | o4 miniCreator not reportedEvaluated 2025-06-30 | 91.6% | Not reported | Unknown |
| 8 | gpt-oss-120bCreator not reportedEvaluated 2025-08-19 | 91.2% | Not reported | Unknown |
| 9 | gpt-oss-20bCreator not reportedEvaluated 2025-08-19 | 91.2% | Not reported | Unknown |
| 10 | Claude 3.7 SonnetCreator not reportedEvaluated 2025-07-01 | 90.4% | Not reported | Unknown |
| 11 | DeepSeek-V3Creator not reportedEvaluated 2025-06-30 | 89.2% | Not reported | Unknown |
| 12 | deepseek-v3.1Creator not reportedEvaluated 2025-08-27 | 89.2% | Not reported | Unknown |
| 13 | Gemini 1.5 ProCreator not reportedEvaluated 2025-06-30 | 88.8% | Not reported | Unknown |
| 14 | GPT-4oCreator not reportedEvaluated 2025-06-30 | 88.8% | Not reported | Unknown |
| 15 | DeepSeek-R1Creator not reportedEvaluated 2025-06-30 | 88.8% | Not reported | Unknown |
| 16 | Claude 3.5 SonnetCreator not reportedEvaluated 2025-06-30 | 88.4% | Not reported | Unknown |
| 17 | o3 miniCreator not reportedEvaluated 2025-06-30 | 88.4% | Not reported | Unknown |
| 18 | Grok 3 MiniCreator not reportedEvaluated 2025-06-30 | 88.4% | Not reported | Unknown |
| 19 | Gemini 3 Pro PreviewCreator not reportedEvaluated 2025-11-14 | 88.4% | Not reported | Unknown |
| 20 | Gemini 1.5 FlashCreator not reportedEvaluated 2025-06-30 | 87.6% | Not reported | Unknown |
| 21 | Gemini 2.5 Flash PreviewCreator not reportedEvaluated 2025-06-30 | 87.6% | Not reported | Unknown |
| 22 | o3Creator not reportedEvaluated 2025-07-01 | 87.6% | Not reported | Unknown |
| 23 | Claude Sonnet 4Creator not reportedEvaluated 2025-07-01 | 85.6% | Not reported | Unknown |
| 24 | Gemini 3.5 FlashCreator not reportedEvaluation date not reported | 83.5% | Not reported | Unknown |
| 25 | Mistral Large 2Creator not reportedEvaluated 2025-07-01 | 83.2% | Not reported | Unknown |
| 26 | Grok 4Creator not reportedEvaluated 2025-09-02 | 82.8% | Not reported | Unknown |
| 27 | GPT-4o miniCreator not reportedEvaluated 2025-06-30 | 82.0% | Not reported | Unknown |
| 28 | Claude 3.5 HaikuCreator not reportedEvaluated 2025-07-01 | 78.8% | Not reported | Unknown |
| 29 | Gemini 1.5 Flash 8BCreator not reportedEvaluated 2025-06-30 | 78.4% | Not reported | Unknown |
| 30 | Grok 3Creator not reportedEvaluated 2025-06-30 | 78.0% | Not reported | Unknown |
| 31 | Gemma 3 12BCreator not reportedEvaluated 2025-06-30 | 74.4% | Not reported | Unknown |
| 32 | CodestralCreator not reportedEvaluated 2025-07-01 | 74.0% | Not reported | Unknown |
| 33 | Gemini 2.0 FlashCreator not reportedEvaluated 2025-06-30 | 72.8% | Not reported | Unknown |
| 34 | o1Creator not reportedEvaluated 2025-06-30 | 66.8% | Not reported | Unknown |
| 35 | Qwen 3 235B A22B ThinkingCreator not reportedEvaluated 2025-08-19 | 65.6% | Not reported | Unknown |
| 36 | Gemma 3 4BCreator not reportedEvaluated 2025-06-30 | 64.8% | Not reported | Unknown |
| 37 | Mixtral 8x22BCreator not reportedEvaluated 2025-07-01 | 58.8% | Not reported | Unknown |
| 38 | Gemma 4 31BCreator not reportedEvaluation date not reported | 57.2% | Not reported | Unknown |
| 39 | Ministral 8BCreator not reportedEvaluated 2025-07-01 | 56.0% | Not reported | Unknown |
| 40 | GPT-4.1Creator not reportedEvaluated 2025-06-30 | 54.0% | Not reported | Unknown |
| 41 | Granite 4.0 SmallCreator not reportedEvaluated 2025-09-30 | 53.2% | Not reported | Unknown |
| 42 | Grok 2Creator not reportedEvaluated 2025-06-30 | 50.4% | Not reported | Unknown |
| 43 | Gemma 3 27BCreator not reportedEvaluated 2025-06-30 | 48.8% | Not reported | Unknown |
| 44 | o1 miniCreator not reportedEvaluated 2025-06-30 | 46.4% | Not reported | Unknown |
| 45 | GPT-3.5 TurboCreator not reportedEvaluated 2025-06-30 | 32.8% | Not reported | Unknown |
| 46 | Ministral 3BCreator not reportedEvaluated 2025-07-01 | 24.8% | Not reported | Unknown |
How to read this leaderboard
Operator receipt: 46 sourced rows are currently displayable on this page; the leading published row is Claude Opus 4.8 at 95.6%.
Honest limit: The public download does not specify prompting and generation settings for each row. Some rows omit the display name; their source slug is shown verbatim. A result in Bengali does not establish performance in every language.
How to read MGSM Bengali
These results cover the Bengali translation of 250 grade-school math questions. The public Kaggle download includes configurations hidden by the website's default model filters.
We retain the published evaluation dates where available. The capture date is when the source was downloaded. Prompting settings remain unknown, and these results do not enter overall or category rankings.
Snapshot
These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. The table retains each source configuration and its reported value.
46 models are shown for MGSM Bengali (Kaggle). The benchmark falls in the Mathematics category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. MGSM Bengali (Kaggle) is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About MGSM Bengali (Kaggle)
Year
2022
Tasks
250 Bengali math problems
Format
Problem accuracy (%)
The Bengali subset has 250 manually translated problems. We show the exact published Kaggle Score and keep its source model profiles and evaluation dates. This single-language protocol is separate from the multilingual MGSM aggregate and the Vals evaluation.
Freshness and provenance
Version
MGSM Bengali (Kaggle) 2022
Refresh cadence
Static
Staleness state
Stale
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does MGSM Bengali (Kaggle) measure?
Tests mathematical reasoning on the Bengali translation of 250 grade-school math problems from GSM8K.
Which source row has the highest reported MGSM Bengali (Kaggle) value?
Claude Opus 4.8 has the highest reported MGSM Bengali (Kaggle) value at 95.6%. These are published source configurations. Compare the task version and disclosed settings before treating a score difference as a model difference. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on MGSM Bengali (Kaggle)?
The 2026-10-20 source capture snapshot contains 46 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.