# MiniCPM

> Source: https://aiwiki.ai/wiki/minicpm
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> 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. "MiniCPM." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/minicpm
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**MiniCPM** is a family of compact, openly licensed language models published by [OpenBMB](https://aiwiki.ai/wiki/openbmb), the shared open-source brand of [Tsinghua University](https://aiwiki.ai/wiki/tsinghua_university)'s natural language processing lab (THUNLP) and the Beijing company ModelBest.[1][2] The line opened on February 1, 2024 with a 2.4-billion-parameter text model that its authors said performed comparably to Mistral-7B on public benchmarks, and by September 2026 it covered five text generations, a vision-language branch (MiniCPM-V), an omni-modal branch (MiniCPM-o), a speech-synthesis branch (VoxCPM), and a set of sparse, quantized and long-context research variants.[1][2][3] Every model in the line is sized to run locally, on a phone, a laptop, a consumer GPU or an embedded board, rather than in a data center.

The framing OpenBMB's own researchers use for what they are optimizing is capacity density. In "Densing Law of LLMs" (arXiv:2412.04315, December 2024), Chaojun Xiao, Xu Han, Zhiyuan Liu, Maosong Sun and colleagues define capacity density as the ratio of a model's effective parameter size, meaning the size a reference model would need in order to reach the same downstream score, to its actual parameter size, and report that across recent open base models this quantity has doubled roughly every three months.[4] That is the group's own metric and its own empirical claim, not an independently replicated finding, but it explains the shape of the MiniCPM roadmap: the flagship has moved down in size at least as often as it has moved up.

## The series at a glance

The tables below list every generation and variant that could be verified against a live Hugging Face model card, the repository changelogs, or a published paper. Release dates are the announcement dates in the `OpenBMB/MiniCPM` and `OpenBMB/MiniCPM-V` changelogs, which sometimes trail the Hub repository creation date by a few weeks.[1][2] Parameter counts marked as vendor figures are the ones stated on the model card; where a safetensors index was readable, the exact tensor total from the Hugging Face API is given instead.[3]

### Text models

| Model | Announced | Parameters | Architecture | Context | License on weights today |
| --- | --- | --- | --- | --- | --- |
| MiniCPM-2B | February 1, 2024 | 2.4B non-embedding | Dense | 4,096 | General Model License (card unchanged)[5] |
| MiniCPM-1B | April 11, 2024 | 1.2B non-embedding | Dense | 4,096 | General Model License (card unchanged)[5] |
| MiniCPM-2B-128k | April 11, 2024 | 2.4B non-embedding | Dense | 128K (65,536 base positions with dynamic RoPE scaling) | General Model License (card unchanged)[5] |
| MiniCPM-MoE-8x2B | April 11, 2024 | 13.6B non-embedding, about 4B active | [Mixture of experts](https://aiwiki.ai/wiki/mixture_of_experts), 2 of 8 experts per token | 4,096 | General Model License (card unchanged)[5] |
| MiniCPM-S-1B | July 5, 2024 | 1.36B total | Dense with ReLU activation sparsity (ProSparse), 87.89% FFN sparsity | 4,096 | Apache 2.0[6] |
| MiniCPM3-4B | September 5, 2024 | 4B (vendor figure) | Dense, low-rank latent attention (`q_lora_rank` 768, `kv_lora_rank` 256) | 32,768 | Tagged Apache 2.0, body text still cites the MiniCPM Model License[7] |
| MiniCPM4-0.5B | June 6, 2025 | 0.43B total | Dense with InfLLM v2 trainable [sparse attention](https://aiwiki.ai/wiki/sparse_attention) | 32,768 native | Apache 2.0[8] |
| MiniCPM4-8B | June 6, 2025 | 8.19B total | Dense with InfLLM v2 | 32,768 native, validated to 131,072 with LongRoPE | Apache 2.0[8] |
| BitCPM4-0.5B / BitCPM4-1B | June 6, 2025 | 0.43B / 1.36B total | Ternary quantization, applied to MiniCPM4-0.5B for the 0.5B model and to MiniCPM3-1B for the 1B model | 32,768 | Apache 2.0[3][9] |
| MiniCPM4.1-8B | September 5, 2025 | 8.19B total | Dense with InfLLM v2, hybrid reasoning | 65,536 native, validated to 131,072 | Apache 2.0[10] |
| MiniCPM-SALA | February 11, 2026 | 9.48B total | Hybrid attention: 25% InfLLM v2 sparse layers, 75% Lightning linear attention | 524,288 configured; 1M-token inference demonstrated | Apache 2.0[11] |
| MiniCPM5-1B | May 19, 2026 | 1,080,632,832 | Dense, standard `LlamaForCausalLM` | 131,072 | Apache 2.0[3] |
| MiniCPM5-2B | September 7, 2026 | 2,516,756,480 | Dense, standard `LlamaForCausalLM` | 131,072 | Apache 2.0[12] |

### Multimodal models

| Model | Announced | Total parameters | Components | Context | Modality |
| --- | --- | --- | --- | --- | --- |
| MiniCPM-V 1.0 (OmniLMM-3B) | February 1, 2024 | 3B on the card, 2.8B in the technical report; 3.43B in the checkpoint | SigLIP-400M plus MiniCPM-2.4B, perceiver resampler, 64 query tokens | 4,096 | Image, text |
| MiniCPM-V 2.0 | April 12, 2024 | 2.8B (vendor); 3.43B in the checkpoint | SigLIP-400M plus MiniCPM-2.4B, LLaVA-UHD adaptive encoding | 4,096 | Image, text |
| MiniCPM-Llama3-V 2.5 | May 20, 2024 | 8B (vendor); 8.54B in the checkpoint | SigLIP-400M plus Llama3-8B-Instruct, 96 tokens per slice | 8,192 | Image, text |
| MiniCPM-V 2.6 | August 6, 2024 | 8B (vendor); 8.10B in the checkpoint | SigLIP-400M plus Qwen2-7B | Repository gated as of September 8, 2026 | Image, multi-image, video, text |
| MiniCPM-o 2.6 | January 13, 2025 | 8B (vendor); 8.67B in the checkpoint | SigLIP-400M, Whisper-medium, ChatTTS, Qwen2.5-7B | 32,768 | Vision, speech, streaming |
| MiniCPM-V 4.0 | August 2, 2025 | 4.1B (vendor); 4.06B in the checkpoint | SigLIP2-400M plus MiniCPM4-3B | 32,768 | Image, video, text |
| MiniCPM-V 4.5 | August 26, 2025 | 8B (vendor); 8.70B in the checkpoint | SigLIP2-400M plus Qwen3-8B, unified 3D-Resampler | 40,960 | Image, multi-image, high-FPS video |
| MiniCPM-o 4.5 | February 3, 2026 | 9B (vendor); 9.37B in the checkpoint | SigLIP2, Whisper-medium, CosyVoice2, Qwen3-8B | 40,960 | Full-duplex omni-modal |
| MiniCPM-V 4.6 and MiniCPM-V 4.6-Thinking | May 11, 2026 | 1.3B | SigLIP2-400M plus Qwen3.5-0.8B, mixed 4x/16x visual token compression | 262,144 | Image, multi-image, video, text |

Every multimodal release carries [Apache 2.0](https://aiwiki.ai/wiki/apache_license) on its code. The weights of MiniCPM-V 1.0, 2.0 and Llama3-V 2.5 are still governed by the older MiniCPM Model License. MiniCPM-o 2.6 and everything from MiniCPM-V 4.0 onward is tagged Apache 2.0 on the Hub. The MiniCPM-V 2.6 repository sits apart from both: it is gated (`gated: auto` in the Hugging Face API as of September 8, 2026, so files require a signed-in account) and carries no license tag.[2][3][13] The vision branch is covered in depth at [MiniCPM-V](https://aiwiki.ai/wiki/minicpm_v).

### Speech models

| Model | Announced | Backbone parameters | Sample rate | Languages | License |
| --- | --- | --- | --- | --- | --- |
| VoxCPM-0.5B | September 2025 | 0.5B | 16 kHz | Chinese, English | Apache 2.0[14][15] |
| VoxCPM1.5 | December 2025 | 0.6B (0.80B in the checkpoint) | 44.1 kHz | Chinese, English | Apache 2.0[14] |
| VoxCPM2 | April 2026 | 2B (2.29B in the checkpoint) | 48 kHz | 30 | Apache 2.0[14][16] |

VoxCPM is a tokenizer-free diffusion autoregressive text-to-speech system rather than a language model with an audio head; VoxCPM2 is built on a MiniCPM-4 backbone and operates in the latent space of an AudioVAE.[14] The repository lists the original VoxCPM technical report with an ICLR 2026 OpenReview link alongside its arXiv entry.[14][15]

## What distinguishes MiniCPM technically

The design argument was set out in the first paper, "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies" (arXiv:2404.06395), by Shengding Hu and 24 co-authors.[17] Three ideas from it recur through every later generation.

The first is the model wind tunnel. Rather than tuning hyper-parameters for each target size, the team ran extensive sweeps on small models and transferred the results upward, using the width and depth scaling rules from the Tensor Programs line of work. The paper names the practice after aircraft wind-tunnel testing and splits it into three parts: hyper-parameter transfer, optimal batch-size scaling, and learning-rate stability.[17]

The second is the WSD learning-rate schedule. Instead of a single cosine decay across a fixed token budget, warmup-stable-decay holds the learning rate flat for a long stable phase and then decays it over a short final phase. The paper reports that loss falls sharply during the decay stage, and that the checkpoint taken just before decay can be reused as the starting point for further training. That makes it possible to add data continuously without committing in advance to a total token count, which a [cosine schedule](https://aiwiki.ai/wiki/cosine_schedule) forces you to do, and it shows up as a distinctive step at the end of the [loss curve](https://aiwiki.ai/wiki/loss_curve).[17] The same property let the team measure a data-model scaling law cheaply: their fit puts the compute-optimal data size at roughly 192 times the model size, against about 20 times in the Chinchilla work.[17]

The third is that efficiency work belongs in the architecture, not only in the serving stack. MiniCPM-S replaced the Swish activation with ReLU under the ProSparse method to reach 87.89 percent activation sparsity in the feed-forward layers.[6][18] MiniCPM4 introduced InfLLM v2, a trainable sparse attention mechanism in which each token attends to under 5 percent of the context in 128K processing, alongside the UltraClean data filter, the ModelTunnel v2 pre-training search, and the ternary BitCPM quantization.[8][19] MiniCPM-SALA then combined sparse and linear attention in one model, putting InfLLM v2 in a quarter of the layers and Lightning Attention in the rest, which its card says allows million-token inference on a single consumer NVIDIA RTX 5090.[11]

InfLLM v2 was later written up as a standalone paper, "InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation" (arXiv:2509.24663, September 2025), which describes the mechanism as reusing the dense attention parameters through a parameter-free architectural change, running dense attention on short inputs and switching to sparse attention on long ones, and reports a 4x speedup over dense attention while retaining 98.1 percent of long-context and 99.7 percent of chain-of-thought reasoning performance in its own experiments.[26] Two later papers built on that backbone.

The first is in-house. [NOSA](https://aiwiki.ai/wiki/nosa) (Native and Offloadable Sparse Attention), from THUNLP and OpenBMB (arXiv:2510.13602, first posted October 15, 2025, revised January 29, 2026), is a trainable sparse attention designed so that most of the [KV cache](https://aiwiki.ai/wiki/kv_cache) can sit in CPU memory during decoding. The paper builds NOSA "on top of InfLLMv2", keeping its two-stage block-wise selection, and splits each step's block budget into a query-aware part and a query-agnostic part chosen by a trainable eviction head, which bounds how many KV blocks must cross the PCIe bus per step. It trains 1B, 3B and 8B models: the 1B and 3B use a Llama-3 architecture, the 8B takes MiniCPM4-base as its base model, and all three share the MiniCPM4 tokenizer. With its NOSI inference system, the authors report decoding throughput up to 5.04x over full attention, 1.92x over InfLLM v2 and 1.83x over ShadowKV, on an NVIDIA A800.[27]

The second is external. In June 2026 NVIDIA researchers used MiniCPM4.1-8B, together with NOSA-8B, as the two sparse-pretrained models on which they evaluated [SparDA](https://aiwiki.ai/wiki/sparda) (Sparse Decoupled Attention, arXiv:2606.04511). SparDA adds a fourth per-layer projection, the Forecast, that predicts which KV blocks the next layer will select, so a serving runtime can prefetch them from CPU memory one layer ahead and the selector can run with one head per grouped-query attention group instead of one per query head. Only the Forecast projections are trained, adding 33.5 million parameters (0.41 percent of the 8B total) with the backbone frozen; for MiniCPM4.1-8B this took under 48 hours on 32 H100 GPUs at 64K sequence length. Accuracy was measured at the model's 64K native maximum, and RULER runs beyond it used the 128K-validated LongRoPE factors shipped with the checkpoint.[28]

| MiniCPM4.1-8B configuration (SparDA paper, Table 1) | HELMET | LongBench | RULER | Reasoning | Average |
| --- | --- | --- | --- | --- | --- |
| Dense attention | 41.7 | 44.8 | 85.3 | 82.3 | 63.5 |
| Sparse (InfLLM v2, top-k selection) | 38.9 | 45.0 | 78.2 | 83.6 | 61.4 |
| InfiniGen (training-free lookahead) | 33.5 | 45.1 | 68.4 | 83.7 | 57.7 |
| SparDA | 38.3 | 45.1 | 78.7 | 84.7 | 61.7 |

The reasoning column is MATH-500, AIME 2024 and AIME 2025 with thinking mode enabled. On efficiency, measured on an H100 at 128K context, the paper reports prefill throughput of 17,087.6 tokens per second for SparDA against 13,661.8 for the InfLLM v2 sparse-offload baseline (1.25x) and 8,085.8 for dense attention (2.11x), decode speedups of up to 1.69x over the sparse-offload baseline, and, because offloading permits larger batches on one GPU, up to 5.28x higher decode throughput than the non-offload sparse configuration. The authors say the gap between dense and sparse attention on HELMET and RULER is likely due to evaluating at 64K when the model's sparse-attention pretraining was conducted at 32K. These are NVIDIA's measurements on its own runtime, and SparDA is an add-on whose accuracy is bounded by the InfLLM v2 backbone it sits on.[28]

MiniCPM5 reversed direction on one point. Where MiniCPM4 built its efficiency case on custom attention and a custom inference stack, the MiniCPM5 cards state that the models use the plain `LlamaForCausalLM` architecture so that mainstream engines can load them without a model-code fork.[12] The published configs bear this out: `model_type` is `llama` for both MiniCPM5-1B and MiniCPM5-2B.[3][12]

Note that arXiv:2506.07900 is the MiniCPM4 technical report, not a MiniCPM5 paper, even though both current MiniCPM5 cards cite it as "MiniCPM Tech Report".[12][19]

## Efficiency and deployment

The line is packaged for local inference to an unusual degree. As of September 8, 2026 the openbmb Hugging Face account carried 172 model repositories, a large share of which are alternate builds of a handful of base models: [GGUF](https://aiwiki.ai/wiki/gguf) for [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) and [Ollama](https://aiwiki.ai/wiki/ollama), [MLX](https://aiwiki.ai/wiki/mlx) for [Apple Silicon](https://aiwiki.ai/wiki/apple_silicon), [GPTQ](https://aiwiki.ai/wiki/gptq), [AWQ](https://aiwiki.ai/wiki/awq), bitsandbytes int4, Marlin kernels, Eagle and DSpark draft models for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding), and CANN builds for Huawei Ascend.[3]

Upstream support has followed each flagship rather than preceding it. MiniCPM-V 4.5 gained official llama.cpp, [vLLM](https://aiwiki.ai/wiki/vllm) and LLaMA-Factory support on September 1, 2025, with [SGLang](https://aiwiki.ai/wiki/sglang) and Ollama following.[2] MiniCPM-V 4.6 was merged into Ollama's official model library on June 25, 2026.[2] MiniCPM4.1 has an official Ollama entry under the `openbmb` namespace.[1] OpenBMB also maintains its own inference paths where the mainstream ones lag: CPM.cu for CUDA end devices and ArkInfer for cross-platform deployment.[19]

The on-device claims are specific and hardware-bound. The MiniCPM-V 4.0 card reports under two seconds to first token and more than 17 tokens per second decoding on an iPhone 16 Pro Max.[20] MiniCPM-V 4.6 is packaged for iOS, Android and HarmonyOS with the edge adaptation code released, and its card says the intra-ViT early compression technique from LLaVA-UHD v4 (arXiv:2605.08985) cuts visual encoding FLOPs by more than half.[13] These are OpenBMB's own measurements on hardware it chose. See [edge AI](https://aiwiki.ai/wiki/edge_ai) and [small language model](https://aiwiki.ai/wiki/small_language_model) for the wider category.

## Reception and independent evaluation

Most numbers on the MiniCPM cards are self-reported, either OpenCompass averages the team ran itself or head-to-head tables against a comparison set the authors picked. Two releases have independent numbers from [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis).

MiniCPM-V 4.6 scored 13 on the Artificial Analysis Intelligence Index in an article dated May 11, 2026, which the firm called the highest for any open-weights model under 2 billion parameters, alongside 38 percent on MMMU-Pro and 5.4 million output tokens to run the index.[21] Artificial Analysis's own version history places Intelligence Index v4.0.4 in force from March to June 2026, so that figure is not comparable with scores published under the v4.2 and v4.3 index in use by September 2026.[22]

[MiniCPM5-2B](https://aiwiki.ai/wiki/minicpm5_2b) drew a fuller writeup on September 7, 2026, and also an object lesson in index versioning: the same model scored 23 on v4.1.1 and 15 on v4.2, which are not comparable to each other. That article covers the dispute in detail.

Adoption is easier to measure than quality. Over the 30 days to September 8, 2026 the Hugging Face API reported 753,030 downloads for MiniCPM-o 4.5, 724,212 for MiniCPM5-1B, 530,314 for MiniCPM-V 4.6, 410,797 for VoxCPM2 and 264,825 for MiniCPM-V 4.5.[3]

## Licensing history

The weight license changed twice, so a single label for the whole family would be wrong.

At launch in February 2024 the repository was Apache 2.0 but the weights were bound to the General Model License, a Chinese license family OpenBMB publishes itself. The README of that period says the weights are "completely free for academic research" and that commercial use requires writing to cpm@modelbest.cn for a certificate of authorization.[5][23] By September 2024 the terms had moved to a bespoke "MiniCPM Model License": still free for research, and free for commercial use after filling in a registration questionnaire.[24] The restriction was removed with the MiniCPM4 release. A commit dated June 6, 2025 with the message "change license" rewrote the section to a single line: "This repository and MiniCPM models are released under the Apache-2.0 License."[25]

The change was applied to current models rather than swept retroactively across the archive. The MiniCPM-2B, MiniCPM-1B, MiniCPM-2B-128k and MiniCPM-MoE-8x2B cards still carry the General Model License text, and the MiniCPM-V 1.0, 2.0 and Llama3-V 2.5 cards still point at the MiniCPM Model License with its questionnaire.[5][13] MiniCPM-o 2.6, released before the switch, has since been relicensed: its card, last modified August 18, 2026, states that the weights and code are Apache 2.0.[13] MiniCPM3-4B is inconsistent with itself, tagged `apache-2.0` in its Hub metadata while its license section still cites the older document.[7] The registration questionnaire survives on the MiniCPM-V repository as an optional request rather than a condition.[2]

## Adjacent lines

Several OpenBMB model families share the MiniCPM base or naming but sit outside the flagship series: MiniCPM-Embedding and MiniCPM-Reranker for retrieval, AgentCPM-GUI, AgentCPM-Explore and AgentCPM-Report for agents, MiniCPM-RobotManip and MiniCPM-RobotTrack for robotics, and the NOSA and InfLLM-V2 research checkpoints described above.[3][27] The organization, its software stack and its dataset releases are covered at [OpenBMB](https://aiwiki.ai/wiki/openbmb).

## See also

- [OpenBMB](https://aiwiki.ai/wiki/openbmb)
- [MiniCPM-V](https://aiwiki.ai/wiki/minicpm_v)
- [MiniCPM5-2B](https://aiwiki.ai/wiki/minicpm5_2b)
- [Small language model](https://aiwiki.ai/wiki/small_language_model)
- [Edge AI](https://aiwiki.ai/wiki/edge_ai)
- [Liu Zhiyuan](https://aiwiki.ai/wiki/liu_zhiyuan)
- [Sparse upcycling](https://aiwiki.ai/wiki/sparse_upcycling)
- [Quantization](https://aiwiki.ai/wiki/quantization)

## References

1. OpenBMB, "MiniCPM" repository README and changelog. Retrieved September 8, 2026. https://github.com/OpenBMB/MiniCPM
2. OpenBMB, "MiniCPM-V" repository README, News section, Model Zoo and license section. Retrieved September 8, 2026. https://github.com/OpenBMB/MiniCPM-V
3. Hugging Face API, model list and per-repository metadata for author `openbmb` (creation dates, license tags, safetensors parameter totals, 30-day download counts). Retrieved September 8, 2026. https://huggingface.co/api/models?author=openbmb
4. Chaojun Xiao, Jie Cai, Weilin Zhao, Guoyang Zeng, Biyuan Lin, Jie Zhou, Zhi Zheng, Xu Han, Zhiyuan Liu, Maosong Sun, "Densing Law of LLMs", arXiv:2412.04315, December 2024. https://arxiv.org/abs/2412.04315
5. OpenBMB, model card `openbmb/MiniCPM-2B-sft-bf16` (license section citing the General Model License). Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM-2B-sft-bf16
6. OpenBMB, model card `openbmb/MiniCPM-S-1B-sft`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM-S-1B-sft
7. OpenBMB, model card and `config.json` for `openbmb/MiniCPM3-4B`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM3-4B
8. OpenBMB, model card and `config.json` for `openbmb/MiniCPM4-8B`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM4-8B
9. OpenBMB, model card `openbmb/BitCPM4-1B`. Retrieved September 8, 2026. https://huggingface.co/openbmb/BitCPM4-1B
10. OpenBMB, model card and `config.json` for `openbmb/MiniCPM4.1-8B`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM4.1-8B
11. OpenBMB, model card and `config.json` for `openbmb/MiniCPM-SALA`, and the MiniCPM-SALA section of the MiniCPM README. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM-SALA
12. OpenBMB, model card and `config.json` for `openbmb/MiniCPM5-2B`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM5-2B
13. OpenBMB, model cards for `MiniCPM-V`, `MiniCPM-V-2`, `MiniCPM-Llama3-V-2_5`, `MiniCPM-o-2_6`, `MiniCPM-V-4`, `MiniCPM-V-4_5`, `MiniCPM-o-4_5`, `MiniCPM-V-4.6` and `MiniCPM-V-4.6-Thinking`, with their `config.json` files. Retrieved September 8, 2026. https://huggingface.co/openbmb
14. OpenBMB, "VoxCPM" repository README, including the VoxCPM2 / VoxCPM1.5 / VoxCPM-0.5B comparison table. Retrieved September 8, 2026. https://github.com/OpenBMB/VoxCPM
15. "VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning", arXiv:2509.24650, September 2025. https://arxiv.org/abs/2509.24650
16. "VoxCPM2 Technical Report", arXiv:2606.06928, June 2026. https://arxiv.org/abs/2606.06928
17. Shengding Hu et al., "MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies", arXiv:2404.06395, April 2024 (v3, June 2024). https://arxiv.org/abs/2404.06395
18. Chenyang Song, Xu Han, Zhengyan Zhang, Shengding Hu, Xiyu Shi, Kuai Li, Chen Chen, Zhiyuan Liu, Guangli Li, Tao Yang, Maosong Sun, "ProSparse: Introducing and Enhancing Intrinsic Activation Sparsity within Large Language Models", arXiv:2402.13516. https://arxiv.org/abs/2402.13516
19. MiniCPM Team, "MiniCPM4: Ultra-Efficient LLMs on End Devices", arXiv:2506.07900, June 2025 (v2, September 2025). https://arxiv.org/abs/2506.07900
20. OpenBMB, model card `openbmb/MiniCPM-V-4`. Retrieved September 8, 2026. https://huggingface.co/openbmb/MiniCPM-V-4
21. Artificial Analysis, "OpenBMB launches MiniCPM-V 4.6 1.3B Instruct", May 11, 2026. https://artificialanalysis.ai/articles/openbmb-launches-minicpm-v-4-6-1-3b-instruct
22. Artificial Analysis, "Intelligence Benchmarking Methodology", including the Intelligence Index version history. Retrieved September 8, 2026. https://artificialanalysis.ai/methodology/intelligence-benchmarking
23. OpenBMB, "MiniCPM" README-en.md at commit b3358343, February 2024. https://github.com/OpenBMB/MiniCPM/blob/b3358343cb6cc40002d92bc382ab92b98d5b8f3e/README-en.md
24. OpenBMB, "MiniCPM" README-en.md at commit c58c9a28, September 28, 2024. https://github.com/OpenBMB/MiniCPM/blob/c58c9a28/README-en.md
25. OpenBMB, "MiniCPM" README-en.md at commit 71245495 ("change license"), June 6, 2025. https://github.com/OpenBMB/MiniCPM/blob/71245495/README-en.md
26. Weilin Zhao, Zihan Zhou, Zhou Su, Chaojun Xiao, Yuxuan Li, Yanghao Li, Yudi Zhang, Weilun Zhao, Zhen Li, Yuxiang Huang, Ao Sun, Xu Han, Zhiyuan Liu, "InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation", arXiv:2509.24663, September 2025. https://arxiv.org/abs/2509.24663
27. Yuxiang Huang, Pengjie Wang, Jicheng Han, Weilin Zhao, Zhou Su, Ao Sun, Hongya Lyu, Hengyu Zhao, Yudong Wang, Chaojun Xiao, Xu Han, Zhiyuan Liu, "NOSA: Native and Offloadable Sparse Attention", arXiv:2510.13602, October 2025 (v2, January 29, 2026). https://arxiv.org/abs/2510.13602
28. Yaosheng Fu, Guangxuan Xiao, Xin Dong, Song Han, Oreste Villa, "SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference", arXiv:2606.04511, June 3, 2026. https://arxiv.org/abs/2606.04511