AlibabaLanguage model
Qwen2-72B-Instruct
Qwen2-72B-Instruct has no overall index value yet because there are too few independent benchmark results.
PopularityRank 493 of 785
32out of 100same as last month
based on search, usage and attention
Summary
Qwen2-72B-Instruct matters chiefly as the starting point for the later Qwen main line. This older dense checkpoint accepts text, follows instructions, and provides 72 billion parameters for direct conversational answers. The repository documents a 131,072-token context window.
Qwen describes the later-generation Qwen 2.5 72B as an update with more knowledge, stronger coding and mathematics, and better structured-data output. It has its own training state and model card.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 81 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.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 67.8
- Score
- 67.8
- Metric
- GSM8k ArithACC
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 67.6
- Score
- 67.6
- Metric
- GSM8k F1
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 52.3
- Score
- 52.3
- Metric
- MATH ArithACC
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 49.2
- Score
- 49.2
- Metric
- MATH F1
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 43.3
- Score
- 43.3
- Metric
- OlympiadBench ArithACC
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 42.1
- Score
- 42.1
- Metric
- OlympiadBench F1
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 39.3
- Score
- 39.3
- Metric
- OmniMATH ArithACC
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 40.2
- Score
- 40.2
- Metric
- OmniMATH F1
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 49.8
- Score
- 49.8
- Metric
- Avg.F1
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 57
- Score
- 57
- Metric
- GSM8k error-final-answer accuracy
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 82.9
- Score
- 82.9
- Metric
- GSM8k correct-final-answer accuracy
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
ProcessBench (Show measurement, test conditions, and source)
- Source value
- 37.7
- Score
- 37.7
- Metric
- MATH error-final-answer accuracy
- Unit
- %
- Category
- mathematical reasoning
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Zhu et al., “Retrieval-Augmented Process Reward Model for Generalizable Mathematical Reasoning,” Findings of ACL 2025
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Table 1 lists this exact model name under “Language Models as Critic”. The authors say these models judge reasoning steps using majority voting. The paper defines ProcessBench F1 as the harmonic mean of accuracies on the incorrect-final-answer and correct-final-answer subsets, and ArithACC as their weighted arithmetic mean. Table 6 gives the distinct accuracies on those two subsets. The paper does not report a complete run configuration or the evaluated weight revision. The captured Hugging Face card revision pins profile identity only, not the weights used in the evaluation.
- Context
- The paper reports no per-model sample count or confidence interval. The evaluated weight revision, precision, serving route, and complete prompt/decoding settings are unpublished and remain unknown. The captured model-card revision identifies the profile only, not the evaluation weights.
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 7, 2026
Qwen2-72B-InstructAmazon Nova Lite 32KAT-Coder-Pro V2.5 32Granite Code 34B 32Claude Haiku 4.5 32
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 Qwen2-72B-Instruct, chosen by model line, developer, size, price and release date.
- Qwen 2.5 72BAlibaba · same developer
- Qwen 2.5 VL 72BAlibaba · same developer
- Qwen2.5-32B-InstructAlibaba · same developer
- Qwen 2.5 Coder 32BAlibaba · same developer
- Qwen 3 Coder NextAlibaba · same developer
- QwQ 32BAlibaba · same developer
- Qwen 3.8 27BAlibaba · same developer
- Qwen 3 14BAlibaba · 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 Qwen2-72B-Instruct
How large is the context window of Qwen2-72B-Instruct?
131,072 tokens. That is how much text Qwen2-72B-Instruct reads at once, about 200 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
Qwen2-72B-InstructRank 94 of 103131K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
Can I run Qwen2-72B-Instruct locally?
Yes, the weights are open.
Stated by Alibaba
- Open weights
- yes
- License
- Qwen License
- Parameters
- 72B
- active per token
- n/a
Sources and data date
Show all 16 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.