OpenBMBLanguage model
MiniCPM5-2B
MiniCPM5-2B has no overall index value yet because there are too few independent benchmark results.
PopularityRank 153 of 785
47out of 100+15 vs. last month
based on search, usage and attention
Summary
MiniCPM5-2B is the larger family member alongside MiniCPM5-1B. The checkpoint has 2.516 billion parameters and the same 131,072-token context. RL and OPD are part of the documented post-training. OpenBMB documents Think as MiniCPM5-2B's only supported mode. The same table lists No-think for MiniCPM5-1B.
The tool-calling format remains XML-based and needs the documented parser in the serving stack for OpenAI-compatible calls. OpenBMB lists No Think only for MiniCPM5-1B.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 71 published benchmark results
Page 1 of 6
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.
CritPt v2: Overall (Show measurement, test conditions, and source)
- Source value
- 0.28571428571428603
- Score
- 0.28571428571428603
- Metric
- pass@1 accuracy
- Unit
- %
- Benchmark version
- v2
- Category
- reasoning
- Task
- overall
- Task label
- Overall
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- disputed
- Retrieved at
- 2026-11-02
- Methodology
- Exact AA configuration: MiniCPM5-2B (release minicpm5-2b, row minicpm5-2b, UUID 641a073a-41c9-4773-9b02-d965c16607c5, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Source model ID: artificial-analysis/641a073a-41c9-4773-9b02-d965c16607c5/minicpm5-2b/minicpm5-2b. Source model name: MiniCPM5-2B. Provider reported by source: OpenBMB. Model snapshot: MiniCPM5-2B. Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Score display: 0.3%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-02-67d46be20a5b. Source payload hash: 67d46be20a5b412db6cfb60ad190441fcaa5513500b8fd47d033168d0e97b6b2. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-11-02. Source cohort: artificial-analysis-critpt-v2:v2:standalone:reasoning:effort=unspecified:suffix=unspecified
- Comparison methodology
- Exact AA configuration: MiniCPM5-2B (release minicpm5-2b, row minicpm5-2b, UUID 641a073a-41c9-4773-9b02-d965c16607c5, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled
- Context
- This AA source field contains the original CritPt score. The earlier CritPt v2 attribution was incorrect; the value and former label remain as disputed historical evidence.. The source dataset marks this result as disputed. It remains visible as documented evidence, but is not reliable comparative proof.
JevBench public-only overall accuracy (Show measurement, test conditions, and source)
- Source value
- 58.44
- Score
- 58.44
- Metric
- Accuracy
- Unit
- %
- Benchmark version
- JevBench bb05a335; dataset dc3995d8; Ardana result page measured 2026-10-13
- Direction
- Higher is better
- Sample
- 231
- Lower confidence bound
- 51.74
- Upper confidence bound
- 65.27
- Source type
- community-reported
- Evaluator
- Ardana JevBench results page
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Ardana evaluator page; JevBench public items only; exact model tag minicpm5:2b; official OpenBMB openbmb/MiniCPM5-2B-GGUF artifact at the revision shown as 2079a22f; default quant MiniCPM5-2B-Q4_K_M.gguf (1.6 GB); stock instruct model; zero-shot chat layout; one run; 95% bootstrap interval resamples items with paraphrase pairs kept together; 600 s client timeout.
- Context
- Ardana is the primary source for its own JevBench run, but institutional independence is not established and attribution is therefore community-reported. The evaluation uses only public JevBench items, including only the public half of the hard tier, and is not comparable with the leaderboard that includes held-out items. One run per quant was performed. Prompt details, sampling parameters and hardware are not stated.
JevBench public-only Easy-tier accuracy (Show measurement, test conditions, and source)
- Source value
- 93.75
- Score
- 93.75
- Metric
- Accuracy; Easy difficulty tier
- Unit
- %
- Benchmark version
- JevBench bb05a335; dataset dc3995d8; Ardana result page measured 2026-10-13
- Direction
- Higher is better
- Sample
- 48
- Source type
- community-reported
- Evaluator
- Ardana JevBench results page
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Ardana evaluator page; JevBench public items only; exact model tag minicpm5:2b; official OpenBMB openbmb/MiniCPM5-2B-GGUF artifact at the revision shown as 2079a22f; default quant MiniCPM5-2B-Q4_K_M.gguf (1.6 GB); stock instruct model; zero-shot chat layout; one run; Difficulty tier Easy; n=48. 600 s client timeout.
- Context
- Ardana is the primary source for its own JevBench run, but institutional independence is not established and attribution is therefore community-reported. The evaluation uses only public JevBench items, including only the public half of the hard tier, and is not comparable with the leaderboard that includes held-out items. One run per quant was performed. Prompt details, sampling parameters and hardware are not stated. No interval is reported for difficulty subsets.
JevBench public-only Standard-tier accuracy (Show measurement, test conditions, and source)
- Source value
- 62.5
- Score
- 62.5
- Metric
- Accuracy; Standard difficulty tier
- Unit
- %
- Benchmark version
- JevBench bb05a335; dataset dc3995d8; Ardana result page measured 2026-10-13
- Direction
- Higher is better
- Sample
- 72
- Source type
- community-reported
- Evaluator
- Ardana JevBench results page
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Ardana evaluator page; JevBench public items only; exact model tag minicpm5:2b; official OpenBMB openbmb/MiniCPM5-2B-GGUF artifact at the revision shown as 2079a22f; default quant MiniCPM5-2B-Q4_K_M.gguf (1.6 GB); stock instruct model; zero-shot chat layout; one run; Difficulty tier Standard; n=72. 600 s client timeout.
- Context
- Ardana is the primary source for its own JevBench run, but institutional independence is not established and attribution is therefore community-reported. The evaluation uses only public JevBench items, including only the public half of the hard tier, and is not comparable with the leaderboard that includes held-out items. One run per quant was performed. Prompt details, sampling parameters and hardware are not stated. No interval is reported for difficulty subsets.
JevBench public-only Hard-tier accuracy (Show measurement, test conditions, and source)
- Source value
- 40.54
- Score
- 40.54
- Metric
- Accuracy; Hard difficulty tier
- Unit
- %
- Benchmark version
- JevBench bb05a335; dataset dc3995d8; Ardana result page measured 2026-10-13
- Direction
- Higher is better
- Sample
- 111
- Source type
- community-reported
- Evaluator
- Ardana JevBench results page
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Ardana evaluator page; JevBench public items only; exact model tag minicpm5:2b; official OpenBMB openbmb/MiniCPM5-2B-GGUF artifact at the revision shown as 2079a22f; default quant MiniCPM5-2B-Q4_K_M.gguf (1.6 GB); stock instruct model; zero-shot chat layout; one run; Difficulty tier Hard; n=111. 600 s client timeout.
- Context
- Ardana is the primary source for its own JevBench run, but institutional independence is not established and attribution is therefore community-reported. The evaluation uses only public JevBench items, including only the public half of the hard tier, and is not comparable with the leaderboard that includes held-out items. One run per quant was performed. Prompt details, sampling parameters and hardware are not stated. No interval is reported for difficulty subsets.
GSM8K test, first 200 questions, BF16 MPS (Show measurement, test conditions, and source)
- Source value
- 86
- Score
- 86
- Metric
- accuracy
- Unit
- %
- Category
- mathematics
- Direction
- Higher is better
- Sample
- 200
- Source type
- community-reported
- Evaluator
- Core AI Model Zoo, MiniCPM5-2B GSM8K first-200 evaluation
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Exact checkpoint openbmb/MiniCPM5-2B in BF16 via Transformers 4.57 on MPS on Mac; the first 200 GSM8K test questions in litertlm-convert evaldata/gsm8k_test.jsonl order; zero-shot CoT suffix, no-think chat template, greedy decoding, maximum 640 new tokens, numeric-normalized answer extraction. Mac hardware model is not stated.
- Context
- Community-reported subset run, not a full standard GSM8K leaderboard result. The source reports no confidence interval or repeated-run uncertainty; do not compare directly with full-test results.
GSM8K test, first 200 questions, 6-bit g8 Neural Engine (Show measurement, test conditions, and source)
- Source value
- 86.5
- Score
- 86.5
- Metric
- accuracy
- Unit
- %
- Category
- mathematics
- Direction
- Higher is better
- Sample
- 200
- Source type
- community-reported
- Evaluator
- Core AI Model Zoo, MiniCPM5-2B GSM8K first-200 evaluation
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. MiniCPM5-2B checkpoint converted to a 6-bit palettized g8 static-graph bundle and run on iPhone 17 Pro, iOS 27.0 (24A437), in the source's TaskEval app; one engine per arm, reset between questions. Evaluation used the first 200 GSM8K test questions in litertlm-convert order, zero-shot CoT suffix, no-think, greedy, maximum 640 new tokens and numeric-normalized extraction. The run stopped at 138 and resumed the next day with the same bundle and container.
- Context
- Community-reported subset run on a quantized device artifact, not a full standard GSM8K result. The source reports no confidence interval. This quantized arm should not be presented as BF16 base-checkpoint performance.
GSM8K test, first 200 questions, 4-bit g32 Neural Engine (Show measurement, test conditions, and source)
- Source value
- 65.5
- Score
- 65.5
- Metric
- accuracy
- Unit
- %
- Category
- mathematics
- Direction
- Higher is better
- Sample
- 200
- Source type
- community-reported
- Evaluator
- Core AI Model Zoo, MiniCPM5-2B GSM8K first-200 evaluation
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Exact MiniCPM5-2B checkpoint in the source's 4-bit g32 Neural Engine bundle on iPhone 17 Pro, iOS 27.0 (24A437), TaskEval app; one engine per arm, reset between questions. Evaluation used the first 200 GSM8K test questions in litertlm-convert order, zero-shot CoT suffix, no-think, greedy, maximum 640 new tokens and numeric-normalized extraction. The source labels this the 2026-09-15 4-bit g32 bundle.
- Context
- Community-reported subset run on a quantized device artifact, not a full standard GSM8K result. The source reports no confidence interval. This quantized arm should not be presented as BF16 base-checkpoint performance.
GSM8K test, first 200 questions, simulated 6-bit g8 on MPS (Show measurement, test conditions, and source)
- Source value
- 86
- Score
- 86
- Metric
- accuracy
- Unit
- %
- Category
- mathematics
- Direction
- Higher is better
- Sample
- 200
- Source type
- community-reported
- Evaluator
- Core AI Model Zoo, MiniCPM5-2B GSM8K first-200 evaluation
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. MiniCPM5-2B weights palettized with coreai-opt's k-means conversion using minicpm5_pal6_g8.yaml, then evaluated with the source's simulation script on MPS on Mac. It is a simulated quantization arm, not the iPhone Neural Engine. Evaluation used the first 200 GSM8K test questions in litertlm-convert order, zero-shot CoT suffix, no-think, greedy, maximum 640 new tokens and numeric-normalized extraction. Mac hardware model is not stated.
- Context
- Community-reported subset run using a simulated quantization route, not a device Neural Engine run or a full standard GSM8K result. The source reports no confidence interval.
GSM8K test, first 200 questions, simulated int8 block-32 on MPS (Show measurement, test conditions, and source)
- Source value
- 85.5
- Score
- 85.5
- Metric
- accuracy
- Unit
- %
- Category
- mathematics
- Direction
- Higher is better
- Sample
- 200
- Source type
- community-reported
- Evaluator
- Core AI Model Zoo, MiniCPM5-2B GSM8K first-200 evaluation
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. MiniCPM5-2B weights evaluated with the source's simulated shipped int8/GPU recipe, symmetric int8 with per-block-32 scales, on MPS on Mac. It is a simulated quantization arm, not the iPhone Neural Engine. Evaluation used the first 200 GSM8K test questions in litertlm-convert order, zero-shot CoT suffix, no-think, greedy, maximum 640 new tokens and numeric-normalized extraction. Mac hardware model is not stated.
- Context
- Community-reported subset run using a simulated quantization route, not a device Neural Engine run or a full standard GSM8K result. The source reports no confidence interval.
GSM8K test, first 100 questions, bf16 PyTorch (MPS), upstream template (Show measurement, test conditions, and source)
- Source value
- 92
- Score
- 92
- Metric
- GSM8K correctness score
- Unit
- %
- Direction
- No direction published
- Sample
- 100
- Source type
- community-reported
- Evaluator
- LiteRT Community, MiniCPM5-2B LiteRT-LM README
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. First 100 GSM8K test questions; greedy decoding; zero-shot chain-of-thought prompt; thinking off; maximum 2048 new tokens; identical prompt and extraction across the table. bf16 PyTorch (MPS), upstream template. The README does not state a confidence interval or the PyTorch runner version.
- Context
- This first-100 subset is nested within the already-recorded first-200 GSM8K test and is not an independent replication.
GSM8K test, first 100 questions, int8 CPU (Show measurement, test conditions, and source)
- Source value
- 91
- Score
- 91
- Metric
- GSM8K correctness score
- Unit
- %
- Direction
- No direction published
- Sample
- 100
- Source type
- community-reported
- Evaluator
- LiteRT Community, MiniCPM5-2B LiteRT-LM README
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. First 100 GSM8K test questions; greedy decoding; zero-shot chain-of-thought prompt; thinking off; maximum 2048 new tokens; identical prompt and extraction across the table. int8 dynamic quantization on linears and embeddings; CPU; LiteRT-LM 0.17.0 as stated for these recommended runtime models. The README does not state a confidence interval.
- Context
- This first-100 subset is nested within the already-recorded first-200 GSM8K test and is not an independent replication.
Page 1 of 6
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
MiniCPM5-2BQwen 2.5 72B 47GPT OSS 120B 47Kimi K2.5 47MiMo V2.6 Distill Qwen 9B 47
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 MiniCPM5-2B, chosen by model line, developer, size, price and release date.
- MiniCPM-V 4.7 1BOpenBMB · same developer
- MiniCPM5-1BOpenBMB · same developer
- MiniCPM-V 4.7 35B-A3BOpenBMB · same developer
- Index-Translate-2BBilibili Index Team · similar size
- Qwen3.5 2BAlibaba · similar size
- Spark-X2.5-1.7BiFLYTEK (SparkLLM Team) · similar size
- Granite 4.2 3BIBM · similar size
- LFM3-VL-2BLiquid AI · 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
- Shutdown
- None announced
- no source
Frequently asked questions about MiniCPM5-2B
How large is the context window of MiniCPM5-2B?
131,072 tokens. That is how much text MiniCPM5-2B reads at once, about 200 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
MiniCPM5-2BRank 94 of 103131K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
Can I run MiniCPM5-2B locally?
Yes, the weights are open.
Stated by OpenBMB
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 2B
- active per token
- n/a
Sources and data date
Show all 18 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 |
|---|---|---|---|
| Artificial Analysis | 10 | 1 | 13 |
Without vendor claims and historical or disputed results; those stay labelled in the single values.