Primary use stated or inferred cautiously from official documentation.
M
MISTRAL / mistralai/LIDstral-Arabic
LIDstral Arabic
Language and dialect classifier publicly released by Mistral on Hugging Face on October 9, 2026. It receives Arabic-script text and distinguishes Modern Standard Arabic, dialects and other languages sharing that writing system. It is neither a generative assistant nor a transcription model. It combines fastText one-vs-all, an MLP classifier and per-class isotonic calibration, returning a label and calibrated probabilities over 51 classes.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | mistralai/LIDstral-Arabic |
|---|---|
| Model type | Text classification |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Local CPU classifier |
| Release | 09 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text |
| Exit | Label and probabilities |
| Context window | Nonempty text after preprocessing |
| maximum output | Label and probabilities over 51 classes |
| Reasoning | Not applicable |
| Published tools | |
| Structured Outings | Local classifier JSON |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · fastText · PyTorch |
Corpus classification, language-based routing, linguistic resource selection and dialect analysis. Including non-Arabic languages reduces confusion between language and script. The official script runs on CPU and offers single inputs, a Python API and JSONL processing. The pipeline requires lid_ova_full.bin, mlp_stacker_ln.pt and isotonic_calibration.pkl to remain together.
It always selects one of 51 classes: unsupported languages or mixed-language texts may be mislabeled. Preprocessing removes web and ASR artifacts, diacritics, decorative elongation and repeated punctuation while preserving alef variants. Empty inputs or text entirely removed by preprocessing raise an error. Calibrated confidence does not replace domain validation or a policy for uncertain cases.
03 / CAPABILITIES
What it can do
Orientation
Corpus classification, language-based routing, linguistic resource selection and dialect analysis. Including non-Arabic languages reduces confusion between language and script. The official script runs on CPU and offers single inputs, a Python API and JSONL processing. The pipeline requires lid_ova_full.bin, mlp_stacker_ln.pt and isotonic_calibration.pkl to remain together.
Context
Nonempty text after preprocessing · Label and probabilities over 51 classes
Tools and integration
Not published
Access
Hugging Face · fastText and PyTorch
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
The model card declares Apache 2.0, and the repository is public and ungated at this review. Its example still mentions a private download; we do not interpret that wording as an available commercial API. No hosted endpoint or per-token price has been verified. Self-hosting entails CPU, memory, storage and operational costs; pin the code revision before loading model files.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Mistral reports evaluation on 84,870 Arabic examples from several datasets. It reports F1 of 88.67% for Moroccan Darija. We retain the dialect and metric: this is not global accuracy and does not extend to every class. Inferama has not reproduced the result. An independent evaluation should inspect the confusion matrix, calibration, mixed texts and out-of-distribution behavior.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| LIDstral Arabic (Moroccan Darija) | 88,67 % | F1 | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
LIDstral Arabic
Public repository release. The October 8 creation date is not used as the release date.
07 / SOURCES