Mistral AILanguage model
Mixtral 8x7B
Mixtral 8x7B has no overall index value yet because there are too few independent benchmark results.
PopularityRank 283 of 785
40out of 100same as last month
based on search, usage and attention
Summary
What is distinctive about Mixtral 8x7B is its sparse mixture-of-experts computation. A router selects two of eight experts per token; Mistral gives 46.7 billion total parameters and about 12.9 billion active parameters with a 32,000-token context window. The release supports five languages, coding, and Apache 2.0.
The larger variant Mixtral 8x22B retains Mixtral 8x7B’s router-and-expert design.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 53 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.
LongHealth Task 1 single-patient information retrieval accuracy (Show measurement, test conditions, and source)
- Source value
- 77
- Score
- 77
- Metric
- Mean accuracy over five runs
- Unit
- %
- Benchmark version
- LongHealth
- Direction
- Higher is better
- Sample
- 400
- Source type
- independent-evaluator
- Evaluator
- LongHealth: A Question Answering Benchmark with Long Clinical Documents
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact checkpoint named in the paper: Mixtral-8x7B-Instruct-v0.1. Zero-shot multiple-choice question answering with the same system prompt for all models, one question at a time, and 16,000-token input length. Task 1 shuffles documents for one patient. The table reports mean accuracy over five runs; sampleCount is the 400 benchmark questions, not five repeated runs.
- Context
- The 75%-79% range is the observed range across five runs, not a confidence interval. No CI is reported. The paper does not specify a served API revision or reasoning-effort setting.
LongHealth Task 2 mixed-patient information retrieval accuracy (Show measurement, test conditions, and source)
- Source value
- 71
- Score
- 71
- Metric
- Mean accuracy over five runs
- Unit
- %
- Benchmark version
- LongHealth
- Direction
- Higher is better
- Sample
- 400
- Source type
- independent-evaluator
- Evaluator
- LongHealth: A Question Answering Benchmark with Long Clinical Documents
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact checkpoint named in the paper: Mixtral-8x7B-Instruct-v0.1. Zero-shot multiple-choice question answering with the same system prompt for all models, one question at a time, and 16,000-token input length. Task 2 mixes target-patient documents with unrelated patients' documents. The table reports mean accuracy over five runs; sampleCount is the 400 benchmark questions, not five repeated runs.
- Context
- The 69%-73% range is the observed range across five runs, not a confidence interval. No CI is reported. The paper does not specify a served API revision or reasoning-effort setting.
LongHealth Task 3 unanswerable-information recognition accuracy (Show measurement, test conditions, and source)
- Source value
- 17
- Score
- 17
- Metric
- Mean accuracy over five runs
- Unit
- %
- Benchmark version
- LongHealth
- Direction
- Higher is better
- Sample
- 400
- Source type
- independent-evaluator
- Evaluator
- LongHealth: A Question Answering Benchmark with Long Clinical Documents
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact checkpoint named in the paper: Mixtral-8x7B-Instruct-v0.1. Zero-shot multiple-choice question answering with the same system prompt for all models, one question at a time, and 16,000-token input length. Task 3 tests refusal when the queried patient's information is absent from the supplied documents. The table reports mean accuracy over five runs; sampleCount is the 400 benchmark questions, not five repeated runs.
- Context
- The 15%-18% range is the observed range across five runs, not a confidence interval. No CI is reported. The paper does not specify a served API revision or reasoning-effort setting.
TALES TextWorld, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 0
- Score
- 0
- Metric
- TextWorld game success rate
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- TextWorld
- Direction
- Higher is better
- Sample
- 10
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. This is the TextWorld aggregate across 10 games; sampleCount is unique games, not attempts.
- Context
- The aggregate table reports no CI. It does not publish the exact serving route, immutable model revision, quantization or a date for this model run. Five runs per game are stated; N here counts games, not multiplied attempts. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
TALES TextWorldExpress, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 1.6
- Score
- 1.6
- Metric
- TextWorldExpress game success rate
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- TextWorldExpress
- Direction
- Higher is better
- Sample
- 16
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. This is TextWorldExpress across 16 games; sampleCount is unique games, not attempts.
- Context
- No CI, exact serving route/revision, quantization or run date is published. Five runs per game are stated; sampleCount counts games, not attempts. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
TALES ALFWorld, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 0
- Score
- 0
- Metric
- ALFWorld game success rate
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- ALFWorld
- Direction
- Higher is better
- Sample
- 12
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. This is ALFWorld across 12 games; sampleCount is unique games, not attempts.
- Context
- No CI, exact serving route/revision, quantization or run date is published. Five runs per game are stated; sampleCount counts games, not attempts. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
TALES ScienceWorld, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 4
- Score
- 4
- Metric
- ScienceWorld game success rate
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- ScienceWorld
- Direction
- Higher is better
- Sample
- 30
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. This is ScienceWorld across 30 games; sampleCount is unique games, not attempts.
- Context
- No CI, exact serving route/revision, quantization or run date is published. Five runs per game are stated; sampleCount counts games, not attempts. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
TALES Jericho, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 0.3
- Score
- 0.3
- Metric
- Jericho game success rate
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- Jericho
- Direction
- Higher is better
- Sample
- 54
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. This is Jericho across 54 games; sampleCount is unique games, not attempts.
- Context
- No CI, exact serving route/revision, quantization or run date is published. Five runs per game are stated; sampleCount counts games, not attempts. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
TALES Average, Mixtral-8x7B-Instruct-v0.1 (Show measurement, test conditions, and source)
- Source value
- 1.3
- Score
- 1.3
- Metric
- Average success rate across TALES environments
- Unit
- %
- Benchmark version
- TALES v4, arXiv:2504.14128v4 Table 3
- Category
- overall average
- Direction
- Higher is better
- Sample
- 122
- Source type
- independent-evaluator
- Evaluator
- TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 3 names Mixtral-8x7B-Instruct-v0.1 and reports the Average across five environments containing 122 games in total. Zero-shot minimal prompt, no domain-expert instructions, non-reasoning model, 100-step cap; five runs per game. sampleCount is unique games, not attempts.
- Context
- The paper does not describe a separate pooled micro-average denominator for this cell or publish a CI. The 122 is the suite's unique game count; five runs per game are stated. Serving revision and route remain unpublished. The authors state that initial-release results include only models reaching at least 90% of the total score in the 100-step Simon Says with Memory mode. Separately, they qualitatively inspect transcripts from nine models scoring highly on the hardest Simon Says mode. This nine-model group is the qualitative subset, not the 34-model zero-shot evaluation set over 122 games.
FILBENCH: Cultural Knowledge (Show measurement, test conditions, and source)
- Source value
- 49.88
- Score
- 49.88
- Metric
- Cultural Knowledge accuracy
- Unit
- %
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- FilBench: Can LLMs Understand and Generate Filipino? (EMNLP 2025)
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Independent evaluation. Table 8 value for the exact listed checkpoint. MCF categories use accuracy; Generation uses ROUGE-L. The source gives no per-model sample count.
- Context
- The source prints 49.88±3.67 but does not define the uncertainty; no confidence interval is inferred.
FILBENCH: Classical NLP (Show measurement, test conditions, and source)
- Source value
- 84.19
- Score
- 84.19
- Metric
- Classical NLP accuracy
- Unit
- %
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- FilBench: Can LLMs Understand and Generate Filipino? (EMNLP 2025)
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Independent evaluation. Table 8 value for the exact listed checkpoint. MCF categories use accuracy; Generation uses ROUGE-L. The source gives no per-model sample count.
- Context
- The source prints 84.19±0.29 but does not define the uncertainty; no confidence interval is inferred.
FILBENCH: Reading Comprehension (Show measurement, test conditions, and source)
- Source value
- 60.95
- Score
- 60.95
- Metric
- Reading Comprehension accuracy
- Unit
- %
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- FilBench: Can LLMs Understand and Generate Filipino? (EMNLP 2025)
- Status
- active
- Retrieved at
- 2026-10-20
- Methodology
- Independent evaluation. Table 8 value for the exact listed checkpoint. MCF categories use accuracy; Generation uses ROUGE-L. The source gives no per-model sample count.
- Context
- The source prints 60.95±0.19 but does not define the uncertainty; no confidence interval is inferred.
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
Mixtral 8x7BGPT OSS 20B 41Aya Expanse 8B 40DeepSeek-V3.2 Exp 40Gemma 3n E2B IT 40
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 Mixtral 8x7B, chosen by model line, developer, size, price and release date.
- Codestral-22B-v0.1Mistral AI · same developer
- Codestral (latest)Mistral AI · same developer
- Mixtral 8x22BMistral AI · same developer
- Mistral Large 2407Mistral AI · same developer
- Pixtral 12BMistral AI · same developer
- Mistral 7B v0.3Mistral AI · same developer
- Yi-34B01.AI · similar size
- DeepSeek LLM 67BDeepSeek · similar size
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 Mixtral 8x7B
How large is the context window of Mixtral 8x7B?
32,768 tokens. That is how much text Mixtral 8x7B reads at once, about 50 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
Mixtral 8x7BRank 103 of 10333K
Can I run Mixtral 8x7B locally?
Yes, the weights are open.
Stated by Mistral AI
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 46.7B
- active per token
- 12.9B
Sources and data date
Show all 23 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. |
| Output | Documented modalities. |
| Research status | Research date October 15, 2026. 16 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | checkpoint-revision Note Open LLM Leaderboard identifies the exact model repository/name but omits the model commit SHA. This historical result is not a current immutable-checkpoint rerun. |
| Not published | immutable-bedrock-model-id-and-serving-revision Original wording The paper names Mixtral 8x7B Instruct on AWS Bedrock but does not disclose the Bedrock model ID or immutable serving revision. |
| Not published | reasoning-effort-quantization-and-decoding-settings Original wording Reasoning effort, quantization, temperature, and other decoding settings are not published. |
| Unresolved | table4-precision-recall-f1-aggregation Original wording The method says both micro and macro metrics were computed but Table 4 does not identify which aggregation is shown for Precision, Recall, and F1. |
| Not published | confidence-intervals Original wording No confidence intervals are reported. |
| Additional source | Open LLM Leaderboard v1 evaluation result snapshot (retrieved October 15, 2026) |
| Additional source | TIGER-Lab official MMLU-Pro leaderboard results.csv (retrieved October 15, 2026) |
| Additional source | Analysis of LLM Performance on AWS Bedrock: Receipt-item Categorisation Case Study (Sanchez et al., 2026) (retrieved October 17, 2026) |
| Additional source | LongHealth: A Question Answering Benchmark with Long Clinical Documents (retrieved October 17, 2026) |
| Additional source | TALES, UC San Diego, Microsoft Research Montréal and Johns Hopkins University (retrieved October 17, 2026) |
| Additional source | EvaLooop paper, arXiv:2505.12185v5 (retrieved October 18, 2026) |
| Additional source | Open LLM Leaderboard evaluation dataset card (retrieved October 19, 2026) |
| Additional source | FilBench: Can LLMs Understand and Generate Filipino? (EMNLP 2025) (retrieved October 20, 2026) |
| Additional source | Arena Text leaderboard (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026)12 |
| Additional source | Model metadata (retrieved October 17, 2026)12 |
| Additional source | Model metadata (retrieved October 19, 2026) |
| Additional source | Mixtral 8x7B 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 |