# MiniCPM5-2B

> Source: https://aiwiki.ai/wiki/minicpm5_2b
> Updated: 2026-09-08
> Fact-checked: 2026-09-08
> Categories: AI Models, Chinese AI, Large Language Models, Open Source AI, Small Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "MiniCPM5-2B." aiwiki.ai, 8 Sept 2026. https://aiwiki.ai/wiki/minicpm5_2b
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

MiniCPM5-2B is an open-weights [small language model](https://aiwiki.ai/wiki/small_language_model) published by [OpenBMB](https://aiwiki.ai/wiki/openbmb) on September 7, 2026 under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license). Its model card gives the parameter count as 2,516,756,480, of which 1,981,982,720 sit outside the embedding matrices, and describes a dense decoder built on the standard `LlamaForCausalLM` architecture with a 131,072-token context window.[1][2] It is the second release in the MiniCPM5 series, following MiniCPM5-1B in May 2026, and OpenBMB positions it for local assistants, coding agents, tool-use workflows and other [edge](https://aiwiki.ai/wiki/edge_ai) or resource-constrained settings rather than for server-scale deployment.[1][3]

The release drew attention because [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) published an independent evaluation the same day, reporting that MiniCPM5-2B scored 15 on the Artificial Analysis Intelligence Index v4.2, the highest of any open-weights model under 4B total parameters.[4] OpenBMB's own launch post on X cited a score of 23 on the same index without naming a version.[5] Artificial Analysis's article resolves that gap directly: it states that MiniCPM5-2B scored 23 on the earlier Intelligence Index v4.1.1 and that "scores across the two versions are not directly comparable."[4] The index moved again within a day. By September 8 the site was running v4.3, and the MiniCPM5-2B model page listed the model at 14, labeled as an estimate.[6]

## Key facts

| Field | Value |
|---|---|
| Developer | OpenBMB[1] |
| Announced | September 7, 2026[3][6] |
| Series | MiniCPM5, second release after MiniCPM5-1B[1] |
| Total parameters | 2,516,756,480 (card); 2.6B as catalogued by Artificial Analysis[1][6] |
| Non-embedding parameters | 1,981,982,720[1] |
| Architecture | Dense causal decoder, standard `LlamaForCausalLM`[1][2] |
| Layers | 42[1][2] |
| Attention | [Grouped-query attention](https://aiwiki.ai/wiki/grouped_query_attention), 16 query heads and 2 key/value heads, head dimension 128[1][2] |
| Hidden size / FFN size | 2,048 / 6,144[2] |
| Vocabulary | 130,560, untied input and output embeddings[2] |
| Context length | 131,072 tokens[1][2] |
| Precision | BF16 release weights[1][2] |
| Modalities | Text in, text out[6] |
| Languages | English and Chinese (card metadata)[1] |
| Knowledge cutoff | December 31, 2025 (per Artificial Analysis)[6] |
| License | Apache 2.0[1][6] |
| Weights | Hugging Face `openbmb/MiniCPM5-2B`, plus Base, SFT, Midtrain, [GGUF](https://aiwiki.ai/wiki/gguf), [GPTQ](https://aiwiki.ai/wiki/gptq), [MLX](https://aiwiki.ai/wiki/mlx) and DSpark repositories; mirrored on [ModelScope](https://aiwiki.ai/wiki/modelscope)[1] |
| Code | github.com/OpenBMB/MiniCPM[3] |

## Place in the MiniCPM line

MiniCPM is OpenBMB's line of compact language models. The wiki covers the lab and its vision branch in more depth at [MiniCPM-V](https://aiwiki.ai/wiki/minicpm_v), which shares the same on-device design goal. The GitHub changelog for the repository records the line's public milestones: MiniCPM-2B in February 2024, MiniCPM3-4B in September 2024, MiniCPM4 in June 2025, MiniCPM4.1 in September 2025, the sparse-and-linear hybrid attention model MiniCPM-SALA in February 2026, MiniCPM5-1B on May 19, 2026, and MiniCPM5-2B on September 7, 2026.[3]

MiniCPM5-1B is the direct predecessor and the reference point for the recipe. Its card describes a 1,080,632,832-parameter dense model with 24 layers, the same 16-query/2-key-value grouped-query attention shape, the same 131,072-token context length, and a hybrid `<think>` chat template switchable through `enable_thinking`.[7] MiniCPM5-2B is described by OpenBMB as scaling "the same training recipe" to a 2B-class dense model.[1] Artificial Analysis classifies MiniCPM5-2B as a reasoning model and notes on the model page that the page covers the reasoning version.[6] The same catalogue lists MiniCPM5-1B (Reasoning) below MiniCPM5-2B on the same index, with both scores flagged as estimates under v4.3.[6]

One point of housekeeping matters for anyone chasing the literature. The MiniCPM5-2B card's "MiniCPM Tech Report" link and its BibTeX block both point at arXiv:2506.07900, which is the MiniCPM4 paper, "MiniCPM4: Ultra-Efficient LLMs on End Devices," dated June 2025 and covering the 0.5B and 8B MiniCPM4 models plus MiniCPM4.1.[1][8] That paper predates MiniCPM5 and does not describe this model. The technical material specific to MiniCPM5-2B is in the model card and the two papers the card cites for its data and reinforcement learning components, discussed below.

## Architecture

MiniCPM5-2B is deliberately unexotic. The card states that it "uses the standard `LlamaForCausalLM` architecture, so mainstream inference engines can load it directly: no custom kernels, no model-code fork."[1] The published `config.json` matches that description: `model_type` is `llama`, there are 42 hidden layers of hidden size 2,048 with an intermediate size of 6,144, SiLU activations, 16 attention heads and 2 key/value heads with a head dimension of 128, an RMS norm epsilon of 1e-6, a vocabulary of 130,560, and no rope scaling with a RoPE theta of 5,000,000.[2] Input and output embeddings are untied, which is where the roughly 535 million parameters separating the total and non-embedding counts sit.[1][2]

That choice is the opposite of the direction MiniCPM4 took, which built its efficiency story on the trainable sparse attention mechanism InfLLM v2 and a custom inference stack.[8] MiniCPM5-2B instead buys compatibility: [vLLM](https://aiwiki.ai/wiki/vllm) 0.21 or later and [SGLang](https://aiwiki.ai/wiki/sglang) 0.5.16 or later serve it directly, and the card's [Transformers](https://aiwiki.ai/wiki/transformers_library) example requires only version 5.6 or later.[1] For tool calling the card recommends SGLang, because the model emits XML-style tool calls and SGLang ships a built-in `minicpm5` parser that converts them into OpenAI-compatible `tool_calls`.[1]

## Training recipe

OpenBMB describes the training of MiniCPM5-2B as "a full-stack practice of UltraData Tiered Data Management," the framework set out in arXiv:2602.09003, "Data Science and Technology Towards AGI Part I: Tiered Data Management," a February 2026 paper from a Tsinghua-affiliated author list including Chaojun Xiao, Xu Han, [Liu Zhiyuan](https://aiwiki.ai/wiki/liu_zhiyuan) and Maosong Sun.[1][9] That paper proposes an L0 to L4 tiered data framework running from raw uncurated resources to organized, verifiable knowledge, with each tier assigned to a different stage of the training lifecycle.[9]

The card breaks the pipeline into three stages:

- **Base training.** Stable training followed by decay training, to build core language capability and training stability.[1]
- **Mid-training.** A further stage to strengthen target capabilities and adapt to the target data distribution. OpenBMB released a `MiniCPM5-2B-Midtrain` checkpoint that captures the model at this point, before supervised fine-tuning.[1]
- **Post-training.** Three steps: [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning), [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning), and on-policy distillation.[1]

The post-training numbers are the most specific claims in the card. OpenBMB states that it used 400B tokens of "deep-thinking SFT" to establish deep-thinking and general chat abilities, then trained specialized RL teachers for math, code, agentic tasks, writing and related domains, and finally used on-policy distillation (OPD) to fold those teachers back into a single release model.[1] For the RL stage the card says it adopted the critic-based algorithm described in "JustRL II," a write-up linked from the card, which follows the JustRL paper (arXiv:2512.16649) from an author list overlapping the MiniCPM team.[1][10][11]

The distillation step is described in some detail. According to the card, OPD merges the capabilities of 16 expert models produced by RL training, five of them agentic experts. At each response position the team computes the full-vocabulary reverse KL divergence between student and teacher logits and uses it as the advantage estimate, replacing the verification-based advantage used during RL, and it reuses the prompts that trained each RL teacher as the distillation data so that no additional corpus has to be built.[1] OpenBMB reports that RL plus OPD improved reasoning and general capabilities by an average of 10.96 points and agentic capabilities by 6.96 points on the benchmarks in its own table.[1] Those are vendor-reported deltas against OpenBMB's own SFT checkpoint, which is itself published as `MiniCPM5-2B-SFT` for anyone who wants to check the comparison.[1]

## Open data and the open recipe

The headline OpenBMB attached to the launch was not only the model but the [open-source](https://aiwiki.ai/wiki/open_source_ai) release of what went into it. Its post said, "We're opening up the data, training recipes, and RL stack behind MiniCPM5-2B."[5] The claim is checkable against the Hugging Face Hub, and the datasets the card names do exist there, all under Apache 2.0:

| Dataset | Role per the model card | On the Hub |
|---|---|---|
| `openbmb/Ultra-FineWeb` | Pre-training corpus | Created March 2025 |
| `openbmb/Ultra-FineWeb-L3` | Pre-training corpus | Created February 2026 |
| `openbmb/UltraX-Preview` | High-quality web pre-training data | Created June 2026 |
| `openbmb/UltraData-Math` | Pre-training and mid-training math data | Created January 2026 |
| `openbmb/UltraData-Code` | Tiered code data, L0 to L3 | Created September 5, 2026 |
| `openbmb/UltraData-SFT-2605` | SFT data | Created May 2026 |
| `openbmb/UltraData-SFT-Agent-2609` | Agent SFT data, described by the card as 500K samples | Created August 31, 2026 |
| `openbmb/UltraData-RL-2609` | RL data, described by the card as 80K+ samples covering mathematics, code, general knowledge and long-context reasoning | Created August 31, 2026 |

Sources for the table: the model card's dataset declarations and highlights, and the Hugging Face dataset metadata for each repository.[1][12] The Hub's own size categories are consistent with the card's counts, listing UltraData-SFT-Agent-2609 in the 100K to 1M range and UltraData-RL-2609 in the 10K to 100K range.[12] The datasets are collected under an UltraData site that OpenBMB describes as a data governance platform aimed at AGI.[1][13]

The "open recipe" side of the claim is partly documentation rather than code. What is published alongside the weights is the four-checkpoint ladder (Base, Midtrain, SFT and the final RL-plus-OPD release), the dataset family above, a prose description of the pipeline in the card, and cookbooks and agent skills in the GitHub repository for deployment and fine-tuning.[1][3] Readers should treat the training-recipe description as OpenBMB's own account of its process rather than as an independently reproduced pipeline.

## Vendor-reported benchmarks

OpenBMB's launch post says that "across 34 benchmarks, MiniCPM5-2B achieves an average score of 53.9," covering coding, math, long-context understanding, tool use and agentic tasks.[5] The model card carries the full table, comparing MiniCPM5-2B with three 2B-class models (LFM2.5-2.6B, [Qwen3.5](https://aiwiki.ai/wiki/qwen3_5)-2B and [Gemma 4](https://aiwiki.ai/wiki/gemma_4)-E2B-it) and five larger ones (Qwen3.5-4B, [granite-4.2-3B](https://aiwiki.ai/wiki/granite_4), [Nemotron-3](https://aiwiki.ai/wiki/nemotron_3)-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B).[1] The card's own framing is that the model "reaches 2B-class open-source SOTA" within that comparison set, with an average of 53.9 against a best-of-the-larger-models figure of 51.1, which belongs to Qwen3.5-4B.[1] The remaining 4B-class averages are 42.7 for granite-4.2-3B, 32.6 for Nemotron-3-Nano-4B, 31.2 for Gemma-4-E4B-it and 28.4 for LFM2.5-8B-A1B.[1]

A footnote on the card matters for reading the table: scores marked with a dagger "come from the official Artificial Analysis release; all others are reproduced internally."[1] Rows in which every cell reproduced below carries the dagger are labelled "AA-sourced"; three rows are mixed and are labelled "partly AA-sourced," with the split given after the table.

| Benchmark | MiniCPM5-2B | LFM2.5-2.6B | Qwen3.5-2B | Gemma-4-E2B-it | Qwen3.5-4B | granite-4.2-3B |
|---|---|---|---|---|---|---|
| Average (34 benchmarks) | 53.9 | 33.2 | 28.0 | 24.6 | 51.1 | 42.7 |
| [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench) v6 | 69.1 | 42.1 | 20.2 | 42.9 | 56.4 | 58.9 |
| LCB-Pro 25Q2 (Easy) | 68.0 | 30.9 | 10.3 | 27.1 | 58.3 | 54.6 |
| LCB-Pro 25Q2 (Medium) | 17.5 | 0.0 | 0.0 | 0.0 | 7.0 | 5.3 |
| OJBench | 32.5 | 11.2 | 2.6 | 11.6 | 24.8 | 21.8 |
| [SciCode](https://aiwiki.ai/wiki/scicode) (wbg), AA-sourced | 26.3 | 14.2 | 2.8 | 20.9 | 16.1 | 24.9 |
| [AIME 2025](https://aiwiki.ai/wiki/aime_2025) | 86.5 | 41.9 | 29.6 | 31.7 | 78.8 | 79.4 |
| AIME 2026 | 86.5 | 45.2 | 29.0 | 39.8 | 82.7 | 83.5 |
| HMMT Feb 2026 | 63.8 | 33.7 | 20.5 | 17.8 | 64.0 | 60.8 |
| [MATH-500](https://aiwiki.ai/wiki/math_500) | 94.6 | 89.6 | 85.8 | 85.4 | 99.0 | 97.0 |
| IFBench | 66.3 | 59.0 | 46.0 | 25.7 | 59.0 | 73.0 |
| [IFEval](https://aiwiki.ai/wiki/ifeval) | 86.7 | 93.4 | 77.5 | 31.4 | 90.2 | 93.7 |
| Multi-IF | 71.8 | 76.8 | 57.1 | 40.3 | 73.6 | 75.9 |
| [MMLU-Pro](https://aiwiki.ai/wiki/mmlu-pro) | 70.8 | 65.2 | 64.3 | 56.0 | 78.0 | 65.8 |
| [MMLU-Redux](https://aiwiki.ai/wiki/mmlu_redux) | 84.7 | 80.0 | 80.0 | 71.8 | 88.7 | 78.9 |
| [HLE](https://aiwiki.ai/wiki/humanity_s_last_exam), AA-sourced | 8.9 | 6.2 | 2.6 | 4.8 | 9.9 | 6.6 |
| [GPQA-Diamond](https://aiwiki.ai/wiki/gpqa_diamond), AA-sourced | 70.2 | 55.8 | 45.6 | 43.3 | 77.1 | 55.9 |
| [SuperGPQA](https://aiwiki.ai/wiki/supergpqa) | 40.8 | 26.2 | 38.6 | 30.3 | 52.8 | 39.9 |
| AA-LCR, AA-sourced | 59.0 | 5.3 | 28.7 | 17.0 | 61.0 | 24.3 |
| [NoLiMa](https://aiwiki.ai/wiki/nolima) | 68.1 | 0.7 | 17.1 | 3.9 | 43.5 | 5.1 |
| LongBenchPro | 44.8 | 23.7 | 8.2 | 42.2 | 58.4 | 34.8 |
| [LongBench v2](https://aiwiki.ai/wiki/longbench_v2) | 43.7 | 30.3 | 24.9 | 33.2 | 47.3 | 36.0 |
| [τ³-Bench](https://aiwiki.ai/wiki/tau-bench) Banking, partly AA-sourced | 20.8 | 7.2 | 2.1 | 3.9 | 6.8 | 5.6 |
| τ²-Bench Telecom, partly AA-sourced | 97.1 | 90.4 | 69.0 | 20.8 | 92.1 | 40.9 |
| [BFCL](https://aiwiki.ai/wiki/bfcl) v4 | 66.6 | 61.1 | 43.6 | 36.6 | 56.8 | 52.2 |
| [SWE-bench Verified](https://aiwiki.ai/wiki/swe_bench_verified) | 46.4 | 6.0 | 5.0 | 2.0 | 33.6 | 36.8 |
| [SWE-bench Pro](https://aiwiki.ai/wiki/swe_bench_pro) | 14.4 | 0.6 | 0.8 | 0.0 | 28.2 | 12.3 |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) v2.1, AA-sourced | 8.6 | 4.5 | 3.0 | 0.4 | 25.8 | 13.9 |
| [BrowseComp](https://aiwiki.ai/wiki/browsecomp)-ZH | 43.5 | 9.8 | 18.2 | 4.7 | 39.6 | 21.1 |
| BrowseComp Top100 | 39.7 | 13.7 | 19.3 | 6.0 | 33.3 | 19.0 |
| [GAIA](https://aiwiki.ai/wiki/gaia_benchmark) Text-103 | 88.7 | 49.5 | 47.9 | 30.1 | 78.6 | 57.3 |
| [GDPval](https://aiwiki.ai/wiki/gdpval)-AA v2, partly AA-sourced | 19.6 | 4.5 | 0.0 | 0.0 | 11.7 | 0.0 |
| Claw-Gym | 59.2 | 19.3 | 25.5 | 31.3 | 51.6 | 60.0 |
| WildClaw | 23.9 | 10.2 | 9.2 | 8.9 | 17.0 | 20.0 |
| QwenClaw | 42.9 | 19.3 | 18.2 | 14.5 | 37.1 | 36.4 |

All figures above are from the model card's table.[1] The columns for Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B are omitted here for width; they are in the card. The three mixed rows split like this within the columns shown: on τ³-Bench Banking the dagger covers MiniCPM5-2B, LFM2.5-2.6B, Qwen3.5-4B and granite-4.2-3B but not Qwen3.5-2B or Gemma-4-E2B-it; on τ²-Bench Telecom it covers Qwen3.5-2B, Gemma-4-E2B-it and Qwen3.5-4B but not MiniCPM5-2B, LFM2.5-2.6B or granite-4.2-3B; on GDPval-AA v2 it covers only MiniCPM5-2B and granite-4.2-3B.[1] The "SOTA" framing is OpenBMB's, applies only to the comparison set the card chose, and is not an independent finding.

## Independent evaluation by Artificial Analysis

Artificial Analysis published its own results on September 7, 2026, the day of the release.[4] Its headline was that MiniCPM5-2B scored 15 on the Artificial Analysis Intelligence Index v4.2, "the highest of any open weights model under 4B total parameters," one point behind Ling 3.0 Tiny (16), a model with roughly three times the total parameters, and four points clear of the next best sub-4B model, Granite 4.2 3B (11).[4] Against larger models the firm reported MiniCPM5-2B one point ahead of Qwen3.5 4B (Reasoning), which it scores at 14 and marks as an estimate, and level with Qwen3.5 9B (Reasoning) at 15, also marked estimated.[4] Artificial Analysis noted that as a dense model, its size advantage is "in memory footprint rather than active-parameter compute."[4]

The per-evaluation results Artificial Analysis reported are the same numbers OpenBMB carried into its own table under the dagger footnote:

| Evaluation (Intelligence Index v4.2 run) | MiniCPM5-2B | Artificial Analysis's comparison note |
|---|---|---|
| GDPval-AA v2 | Elo 831 | Leads models under 4B; human expert baseline is 1,000 |
| τ³-Banking | 21% | Joint first with [Ling 3.0 Tiny](https://aiwiki.ai/wiki/ling_3_0_tiny); next best is Granite 4.2 8B at 8% |
| AA-Briefcase | Elo 438 | Second in the comparison set, above Granite 4.2 8B (324), below Ling 3.0 Tiny (485) |
| Humanity's Last Exam | 9% | 7th in the set, behind Gemma 4 12B (Reasoning) at 16% |
| Terminal-Bench v2.1 | 9% | 8th in the set, behind Qwen3.5 9B (Reasoning) at 29% |
| SciCode | 26% | Second of five measured models, behind Granite 4.2 8B (31%) |
| AA-LCR v1.1 | 59% | 5th in the set, one point behind Ling 3.0 Tiny (60%) |
| GDP.pdf | 1% of tasks passed outright | Behind gpt-oss-20b (high) at 2% |
| CritPt | 0% | Reported as zero |
| AA-Omniscience Index | -12 | Accuracy 8%, Non-Hallucination Rate 78% |

Figures from the Artificial Analysis article of September 7, 2026.[4] Artificial Analysis's live model page carries slightly different values for the two Elo metrics as of September 8, 2026, listing AA-Briefcase at 433.5 and GDPval-AA v2 at 828.4, which is the kind of drift that follows re-runs and re-anchoring.[6]

### Three different index numbers, and why

Three figures for "the Artificial Analysis Intelligence Index" were in circulation within about 24 hours of the launch. They are not interchangeable, and the version history on Artificial Analysis's methodology page explains why.[14]

| Figure | Index version | Source and date |
|---|---|---|
| 23 | v4.1.1 | Artificial Analysis, stated in its September 7, 2026 article; cited without a version in OpenBMB's launch post the same day[4][5] |
| 15 | v4.2 | Artificial Analysis article and model announcement, September 7, 2026; repeated by OpenBMB in a follow-up post naming the version[4][15] |
| 14, marked as an estimate | v4.3 | Artificial Analysis model page, read September 8, 2026[6] |

Artificial Analysis states the relationship between the first two plainly: "MiniCPM5-2B scored 23 on the recently updated Artificial Analysis Intelligence Index v4.1.1. Its new score of 15 on v4.2 reflects the updated evaluation mix and weightings, and scores across the two versions are not directly comparable."[4] The version history records what changed. Version 4.1.1 ran from August 2026 to September 2026. Version 4.2, announced September 4, 2026, added AA-Briefcase at 15% and GDP.pdf at 10%, removed GPQA Diamond from the index, upgraded AA-LCR to v1.1, and rebalanced the weights.[14] Version 4.3, announced September 7, 2026, replaced τ³-Banking with AutomationBench-AA at 5% and replaced Terminal-Bench v2.1 with Terminal-Bench v4.0.[14] As of September 8, 2026 the index in use is v4.3, which Artificial Analysis describes as a weighted average across four categories, Agents at 30%, Coding at 20%, Scientific Reasoning at 20% and General at 30%, drawn from ten evaluations.[14]

The practical consequence is that any statement of the form "MiniCPM5-2B scores X on the Artificial Analysis Intelligence Index" is meaningless without a version and a date. OpenBMB's launch post gave neither, which is why its 23 and Artificial Analysis's 15 read as a contradiction when they are the same model measured under two different index definitions. OpenBMB's second post an hour later did name the version: "On Intelligence Index v4.2, MiniCPM5-2B scores 15, the highest of any open weights model under 4B total parameters."[15]

OpenBMB's launch post also cited "20 on the Agentic Index."[5] Artificial Analysis defines its Agentic Index as "the weighted average of the three agentic evaluations in the Intelligence Index: AA-Briefcase, GDPval-AA v2 and τ³-Banking."[4] Since τ³-Banking left the Intelligence Index at v4.3, that composition has also changed since the launch.[14]

### Token efficiency and abstention

Two findings in the Artificial Analysis writeup are worth separating from the headline number, because they are about behaviour rather than rank.

The first is token efficiency. Artificial Analysis reported that MiniCPM5-2B used 19,000 output tokens per Intelligence Index task, 11,000 of them reasoning tokens, which it called joint-lowest in the comparison set alongside Granite 4.2 3B. Ling 3.0 Tiny spent 56,000 tokens, roughly three times as many, for one more index point, and Granite 4.2 8B spent 33,000.[4] For a model whose intended deployment is on-device, where every generated token is latency and battery, that is arguably a more useful number than the index score itself.

The second is how the model earns its AA-Omniscience result. Artificial Analysis put its Omniscience Index at -12 and attributed that "by abstaining from answering rather than accuracy": the model attempts only 29% of the questions, which gives it a Non-Hallucination Rate of 78%, while its raw accuracy of 8% is a point below Ling 3.0 Tiny (9%) and half that of Qwen3.5 9B (Reasoning) at 16%. Peers that attempt far more questions are penalized much harder, with Qwen3.5 9B (Reasoning) at -53 and gpt-oss-20b (high) at -63.[4] A negative score on that index means more incorrect answers than correct ones, and the scale runs from -100 to 100.[6]

## Deployment

MiniCPM5-2B ships in more packaging formats than most models of its size. Alongside the BF16 release, OpenBMB published a BF16 SFT-only checkpoint, a mid-training checkpoint, a pre-training-only base checkpoint, GGUF builds for [llama.cpp](https://aiwiki.ai/wiki/llama_cpp), [Ollama](https://aiwiki.ai/wiki/ollama) and [LM Studio](https://aiwiki.ai/wiki/lmstudio), a 4-bit MLX build for [Apple Silicon](https://aiwiki.ai/wiki/apple_silicon), a 4-bit GPTQ [quantization](https://aiwiki.ai/wiki/quantization), and a DSpark draft model.[1] The GGUF repository contains F16, Q4_K_M and Q8_0 files.[16] A LiteRT-LM conversion is published separately by the `litert-community` organization on Hugging Face.[1][17]

The DSpark repository is not a variant of the model but a companion for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding). Its card describes it as a draft checkpoint trained for exact pairing with MiniCPM5-2B and its tokenizer, with 5 draft layers, 323,776,001 draft parameters, 7 draft tokens per forward pass, and target hidden-state layers at indices 1, 10, 20, 30 and 39, in BF16.[18] The name refers to the DSpark method described in arXiv:2607.05147, "DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation."[19] The MiniCPM5-2B card shows the SGLang invocation, passing `--speculative-algorithm DSPARK` with the draft model path and a block size of 7, and states that this "accelerates decoding while keeping the target model's outputs unchanged."[1]

The GitHub repository carries per-backend cookbooks and matching agent skills for Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, ArcLight and vLLM Ascend, plus fine-tuning guides for TRL with PEFT, LLaMA-Factory, ms-swift and unsloth.[1][3] The card also documents multi-chip support through FlagOS, the open-source AI system software stack initiated by the Beijing Zhiyuan Research Institute ([BAAI](https://aiwiki.ai/wiki/baai)) with chip makers and other partners, and lists FlagRelease builds of MiniCPM5-2B for nine hardware targets: Nvidia, Hygon, Metax, Iluvatar, Zhenwu, Mthreads, Kunlunxin, Ascend and ARM-v9.[1]

## Limitations

The card carries a plain disclaimer rather than a capability-specific limitations section. It states that the model "has no autonomous intent or legal personhood; its outputs are text generated from statistical patterns and may be inaccurate, biased, or offensive, and may be manipulated by carefully crafted prompts ('jailbreaks') into producing unintended content," and that its responses on politics, health, finance and law are not expert-reviewed and should not be treated as professional advice. The weights are provided "AS IS" without warranty, with users expected to configure their own safeguards and to label AI-generated content where required.[1]

The measured weaknesses are clearer in the independent numbers than in the card. Artificial Analysis identified knowledge, coding and long context as "where it gives ground," placing the model seventh in its comparison set on Humanity's Last Exam, eighth on Terminal-Bench v2.1, and at 0% on CritPt, the physics reasoning evaluation.[4] On GDP.pdf, the professional document reasoning evaluation added in v4.2, it passed 1% of tasks outright, behind gpt-oss-20b (high) at 2%.[4] The high Non-Hallucination Rate comes from answering fewer than a third of AA-Omniscience questions, so it reflects a model tuned to abstain rather than one that knows more.[4] The vendor table and the independent evaluation also disagree in emphasis on coding: OpenBMB reports 46.4 on SWE-bench Verified from an internal reproduction, while the Artificial-Analysis-sourced Terminal-Bench v2.1 figure in the same table is 8.6.[1]

One specification discrepancy is worth noting. The model card gives 2,516,756,480 total parameters, while Artificial Analysis catalogues the model at 2.6B and OpenBMB's own follow-up post refers to "2.6B."[1][4][6][15]

## References

1. [openbmb/MiniCPM5-2B model card](https://huggingface.co/openbmb/MiniCPM5-2B), Hugging Face. Retrieved September 8, 2026.
2. [openbmb/MiniCPM5-2B config.json](https://huggingface.co/openbmb/MiniCPM5-2B/raw/main/config.json), Hugging Face. Retrieved September 8, 2026.
3. [OpenBMB/MiniCPM repository README and changelog](https://github.com/OpenBMB/MiniCPM), GitHub. Retrieved September 8, 2026.
4. Artificial Analysis, ["OpenBMB's MiniCPM5-2B scores 15 on the Artificial Analysis Intelligence Index v4.2, the highest of any open weights model under 4B total parameters"](https://artificialanalysis.ai/articles/openbmb-releases-minicpm5-2b), September 7, 2026.
5. OpenBMB, [launch post announcing MiniCPM5-2B](https://x.com/OpenBMB/status/2096970974247956501), X, September 7, 2026.
6. Artificial Analysis, [MiniCPM5-2B model page](https://artificialanalysis.ai/models/minicpm5-2b). Retrieved September 8, 2026.
7. [openbmb/MiniCPM5-1B model card](https://huggingface.co/openbmb/MiniCPM5-1B), Hugging Face. Retrieved September 8, 2026.
8. MiniCPM Team et al., ["MiniCPM4: Ultra-Efficient LLMs on End Devices"](https://arxiv.org/abs/2506.07900), arXiv:2506.07900, June 2025.
9. Yudong Wang et al., ["Data Science and Technology Towards AGI Part I: Tiered Data Management"](https://arxiv.org/abs/2602.09003), arXiv:2602.09003, February 2026.
10. "JustRL II", Notion write-up [linked by the MiniCPM5-2B model card](https://panhaoxuan.notion.site/justrl-ii-scaling-small-llms-to-128k-reasoning-with-a-critic). Retrieved September 8, 2026.
11. Bingxiang He et al., ["JustRL: Scaling a 1.5B LLM with a Simple RL Recipe"](https://arxiv.org/abs/2512.16649), arXiv:2512.16649, December 2025.
12. Hugging Face dataset repositories for [Ultra-FineWeb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb), [Ultra-FineWeb-L3](https://huggingface.co/datasets/openbmb/Ultra-FineWeb-L3), [UltraX-Preview](https://huggingface.co/datasets/openbmb/UltraX-Preview), [UltraData-Math](https://huggingface.co/datasets/openbmb/UltraData-Math), [UltraData-Code](https://huggingface.co/datasets/openbmb/UltraData-Code), [UltraData-SFT-2605](https://huggingface.co/datasets/openbmb/UltraData-SFT-2605), [UltraData-SFT-Agent-2609](https://huggingface.co/datasets/openbmb/UltraData-SFT-Agent-2609) and [UltraData-RL-2609](https://huggingface.co/datasets/openbmb/UltraData-RL-2609). Retrieved September 8, 2026.
13. [UltraData](https://ultradata.openbmb.cn/), OpenBMB. Retrieved September 8, 2026.
14. Artificial Analysis, [Intelligence Benchmarking Methodology, including Intelligence Index version history](https://artificialanalysis.ai/methodology/intelligence-benchmarking). Retrieved September 8, 2026.
15. OpenBMB, [follow-up post citing Intelligence Index v4.2](https://x.com/OpenBMB/status/2096987494822887716), X, September 7, 2026.
16. [openbmb/MiniCPM5-2B-GGUF](https://huggingface.co/openbmb/MiniCPM5-2B-GGUF), Hugging Face. Retrieved September 8, 2026.
17. [litert-community/MiniCPM5-2B](https://huggingface.co/litert-community/MiniCPM5-2B), Hugging Face. Retrieved September 8, 2026.
18. [openbmb/MiniCPM5-2B-DSpark model card](https://huggingface.co/openbmb/MiniCPM5-2B-DSpark), Hugging Face. Retrieved September 8, 2026.
19. ["DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation"](https://arxiv.org/abs/2607.05147), arXiv:2607.05147.