Mistral AILanguage model
Voxtral Mini 3B 2507
Voxtral Mini 3B 2507 has no overall index value yet because there are too few independent benchmark results.
39out of 100same as last month
based on search, usage and attention
Summary
What is distinctive about Voxtral Mini 3B 2507 is audio understanding in a 3B checkpoint. It accepts audio and text with 32,000 tokens of context, transcribes recordings, answers questions, and creates summaries. It is the 3B variant alongside Voxtral Small 24B.
The model card lists a dedicated transcription mode alongside questions and summaries directly from audio.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 77 published benchmark results
Page 1 of 7
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.
Voxtral Speech Understanding LLM Judge (Show measurement, test conditions, and source)
- Source value
- 85.59
- Score
- 85.59
- Metric
- mean binary helpfulness indicator
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 2, internal Speech Understanding benchmark
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact repository mistralai/Voxtral-Mini-3B-2507, identified with the paper's public Voxtral Mini Online DPO checkpoint. The internal Speech Understanding benchmark uses in-the-wild audio with question-answer prompts. Table 2 reports the mean of a binary correctness-and-helpfulness judge; each answer is independently judged ten times. The total number of questions and full decoding settings are not stated.
- Context
- The table reports 85.59 ± 3.77 as mean ± standard deviation over ten independent judgments per answer. The ± value is not a 95% confidence interval, and the number of evaluated questions is not stated.
Voxtral English Short-Form WER (Show measurement, test conditions, and source)
- Source value
- 6.79
- Score
- 6.79
- Metric
- macro-average word error rate across English short-form test sets
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 2 English short-form evaluation
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact repository mistralai/Voxtral-Mini-3B-2507, identified with the paper's public Voxtral Mini Online DPO checkpoint. Table 2 reports macro-average WER across English short-form test sets. The table does not identify its full dataset list, decoding settings, or number of examples.
- Context
- The paper identifies this as English short-form macro-average WER but Table 2 does not enumerate the datasets or publish an uncertainty interval.
LibriSpeech Test Clean WER (LS-C) (Show measurement, test conditions, and source)
- Source value
- 1.86
- Score
- 1.86
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
LibriSpeech Test Other WER (LS-O) (Show measurement, test conditions, and source)
- Source value
- 4.04
- Score
- 4.04
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
GigaSpeech WER (GS) (Show measurement, test conditions, and source)
- Source value
- 10.68
- Score
- 10.68
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
VoxPopuli WER (VP) (Show measurement, test conditions, and source)
- Source value
- 6.85
- Score
- 6.85
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
SwitchBoard WER (SB) (Show measurement, test conditions, and source)
- Source value
- 11.32
- Score
- 11.32
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
CHiME-4 WER (C-4) (Show measurement, test conditions, and source)
- Source value
- 10.59
- Score
- 10.59
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
SPGISpeech WER (SPGI) (Show measurement, test conditions, and source)
- Source value
- 2.19
- Score
- 2.19
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, English short-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for English short-form WER, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 3 gives no per-corpus sample counts or uncertainty intervals.
Earnings-21 WER (E21 10m) (Show measurement, test conditions, and source)
- Source value
- 9.62
- Score
- 9.62
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, Earnings-21 10m long-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for the Earnings-21/Earnings-22 10m long-form variants, not the separate Voxtral Mini Transcribe or Voxtral Small rows. The paper says these are 10-minute segments of one-hour earnings calls; per-row sample counts and uncertainty intervals are not stated.
Earnings-22 WER (E22 10m) (Show measurement, test conditions, and source)
- Source value
- 11.84
- Score
- 11.84
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 3, Earnings-22 10m long-form
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 3
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is the Voxtral Mini row in Table 3 for the Earnings-21/Earnings-22 10m long-form variants, not the separate Voxtral Mini Transcribe or Voxtral Small rows. The paper says these are 10-minute segments of one-hour earnings calls; per-row sample counts and uncertainty intervals are not stated.
FLEURS Arabic WER (ar) (Show measurement, test conditions, and source)
- Source value
- 25.4
- Score
- 25.4
- Metric
- word error rate (WER)
- Unit
- %
- Benchmark version
- Voxtral paper v1 Table 4, FLEURS per-language
- Category
- audio
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- Mistral AI Voxtral paper Table 4
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact public checkpoint mistralai/Voxtral-Mini-3B-2507. The Voxtral paper identifies Mini Online DPO as the public Voxtral Mini checkpoint (Table 2); this is only the Voxtral Mini row in Table 4, the FLEURS per-language WER table, not the separate Voxtral Mini Transcribe or Voxtral Small rows. Table 4 gives no per-language sample counts or uncertainty intervals.
Page 1 of 7
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 8, 2026
Voxtral Mini 3B 2507Fugu Ultra v2 38Kimi K2 0905 38Kimi K2.8 Preview 38Ling 3.1 Flash 1019 38
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
- 2November 1, 2026: Method version 4 (common scale, shrinkage)
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 Voxtral Mini 3B 2507, chosen by model line, developer, size, price and release date.
- Ministral 3 3BMistral AI · same developer
- Ministral 3 8BMistral AI · same developer
- Ministral 3 14BMistral AI · same developer
- Magistral SmallMistral AI · same developer
- Devstral SmallMistral AI · same developer
- Voxtral Small 24BMistral AI · same developer
- Magistral Small 1.2Mistral AI · same developer
- Voxtral Small (latest)Mistral 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 Voxtral Mini 3B 2507
How large is the context window of Voxtral Mini 3B 2507?
32,768 tokens. That is how much text Voxtral Mini 3B 2507 reads at once, about 50 A4 pages.
Page count as a rule of thumb from Artificial Analysis
Can I run Voxtral Mini 3B 2507 locally?
Yes, the weights are open.
Stated by Mistral AI
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 3B
- active per token
- n/a
Sources and data date
Show all 25 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. 17 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | benchmark-configuration Note Exact numeric corpus and language results are recorded from the Voxtral paper. Not every per-task decoding setting or immutable inference-environment revision is documented. |
| Unresolved | accepted-independent-asr-leaderboard-results Note An exact-checkpoint Open ASR manifest was visible in an HF Staff PR, but it still showed Ready to merge on 2026-10-15. Its proposed values are not treated as accepted leaderboard observations. |
| Not published | test-sample-count-and-decoding-settings Note The evaluator publishes WER and a 95% bootstrap interval but not the number of scored test clips or model decoding settings. |
| Unresolved | accepted-open-asr-result Note The exact Open ASR result is visible as WER 7.05, but the HF result metadata says verified=false and official PR #47 remains open. It is withheld until the benchmark accepts the result. |
| Not published | longshot-model-denominator-and-score-ci Note The paper reports benchmark-wide sample totals and a single generation pass, but no exact model-specific N or score confidence interval. |
| Not published | per-cell-sample-count Original wording The source gives 2,067 clips for the overall robustness domain, but not per subset/model cell. |
| Not published | confidence-intervals Original wording The source reports no confidence interval for Table 7 cells. |
| Unresolved | inference-route Original wording The generic inference section mentions local runs and default settings, while the model appendix names a helper; it does not bind the exact Voxtral row to a route or provide its per-model decoding settings. |
| Additional source | Hugging Face (retrieved October 15, 2026; October 17, 2026; October 22, 2026) · Voxtral Mini 3B model card · Editorial description reviewed October 2, 2026 |
| Additional source | Mistral AI Voxtral paper (retrieved October 15, 2026) |
| Additional source | PERIT-STT (retrieved October 17, 2026) |
| Additional source | LongShOTBench paper authors, arXiv (retrieved October 17, 2026) |
| Additional source | Voice Arena and Hugging Face (retrieved October 17, 2026) |
| Additional source | Handy models (retrieved October 17, 2026) |
| Additional source | AfriSpeech-MultiBench: A Verticalized Multidomain Multicountry Benchmark Suite for African Accented English ASR (retrieved October 18, 2026) |
| Additional source | Efficiently Adapting Spoken Language Models for the Singaporean Context (retrieved October 19, 2026) |
| Additional source | Soniqo, Voxtral Mini 3B 2507 benchmark documentation (retrieved October 19, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026) · Voxtral announcement · Editorial description reviewed October 2, 2026 |
| Additional source | Model metadata (retrieved October 15, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)123456 |
| 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 |