# MiniCPM-o-2_6 by OpenBMB: Open-Weight Model Source: https://savrn.com/models/minicpm-o-2-6 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve MiniCPM-o-2_6 (8.7B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 17.3 GB | 20.8 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 8-bit | 8.7 GB | 10.4 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | | 4-bit | 4.3 GB | 5.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 24, 2026. [MiniCPM-o-2_6 on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/minicpm-o-2-6/gpus) ## SAVRN's Notes on MiniCPM-o-2_6 Plan around 20.8 GB. That is the 16-bit footprint of MiniCPM-o 2.6, and it falls to 10.4 GB at 8-bit and 5.2 GB at 4-bit. On the cheapest card in our table, one MI300X with 192 GB at $1.85 per hour on-demand, that leaves most of the memory free, so the practical deployment is several copies per card. You get an 8.7B parameter any-to-any model assembled end to end from SigLip-400M, Whisper-medium-300M, ChatTTS-200M and Qwen2.5-7B, with a 32,768 token context. Apache 2.0 allows commercial use, modification and redistribution, asks you to keep the license, copyright and NOTICE files and state significant changes, and carries a patent grant from contributors, which matters once a product depends on it. Before committing, check that the alignment data is the RLAIF-V-Dataset, that the weights ship only as safetensors, and that the repository was last updated August 18, 2026 after a January 12, 2025 release. SAVRN Research, 2026-09-18 ## Model Card By OpenBMB, published under apache-2.0, revision 06849bfd36da. ### A GPT-4o Level MLLM for Vision, Speech and Multimodal Live Streaming on Your Phone [GitHub](https://github.com/OpenBMB/MiniCPM-o) | [MiniCPM Wiki(Chinese)](https://modelbest.feishu.cn/wiki/UtWxwcERfiRIpIkBOjuc3h9tn1D) | [Online Demo](https://minicpm-omni-webdemo-us.modelbest.cn) | [Technical Blog](https://openbmb.notion.site/MiniCPM-o-2-6-A-GPT-4o-Level-MLLM-for-Vision-Speech-and-Multimodal-Live-Streaming-on-Your-Phone-185ede1b7a558042b5d5e45e6b237da9) | [Join Us](https://mp.weixin.qq.com/mp/wappoc_appmsgcaptcha?poc_token=HAV8UWijqB3ImPSXecZHlOns7NRgpQw9y9EI2_fE&target_url=https%3A%2F%2Fmp.weixin.qq.com%2Fs%2FKIhH2nCURBXuFXAtYRpuXg%3F) #### News - [2025.06.20] Our official[ollama repository](https://ollama.com/openbmb) is released. Try our latest models with [one click](https://ollama.com/openbmb/minicpm-o2.6)！ - [2025.03.01] RLAIF-V, which is the alignment technique of MiniCPM-o, is accepted by CVPR 2025！The[code](https://github.com/RLHF-V/RLAIF-V), [dataset](https://huggingface.co/datasets/openbmb/RLAIF-V-Dataset), [paper](https://arxiv.org/abs/2405.17220) are open-sourced! - [2025.01.24] MiniCPM-o 2.6 technical report is released![See Here](https://openbmb.notion.site/MiniCPM-o-2-6-A-GPT-4o-Level-MLLM-for-Vision-Speech-and-Multimodal-Live-Streaming-on-Your-Phone-185ede1b7a558042b5d5e45e6b237da9). - [2025.01.19] MiniCPM-o tops GitHub Trending and reaches top-2 on Hugging Face Trending! ### MiniCPM-o 2.6 MiniCPM-o 2.6 is the latest and most capable model in the MiniCPM-o series. The model is built in an end-to-end fashion based on SigLip-400M, Whisper-medium-300M, ChatTTS-200M, and Qwen2.5-7B with a total of 8B parameters. It exhibits a significant performance improvement over MiniCPM-V 2.6, and introduces new features for real-time speech conversation and multimodal live streaming. Notable features of MiniCPM-o 2.6 include: [Read the full model card (3,394 words)](https://savrn.com/models/minicpm-o-2-6/card) ## Configuration Architecture MiniCPMO Context length (tokens) 32,768 Layers 28 Hidden size 3,584 Feed-forward size 18,944 Attention heads 28 Key/value heads 4 Vocabulary size 151,700 Sliding window (tokens) 131,072 RoPE base 1e+06 Stored precision bfloat16 Model type minicpmo ## Identity and Version Repository openbmb/MiniCPM-o-2_6 Publisher OpenBMB Task Any to any Modality Multimodal Library transformers Parameters 8.7B parameters Languages ocr, asr, tts Revision 06849bfd36da94da1f5a88fa17e8ba63f08a4c4d First published 2025-01-12 Last updated 2026-08-18 ## Files and Weights 40 files, 17.4 GB in total. The weights are 5 files totalling 17.4 GB in pt, safetensors. Weights5 files · 17.4 GB Configuration14 files · 424.9 KB Tokenizer6 files · 12.0 MB Documentation1 file · 50.3 KB Other13 files · 15.7 MB Repository1 file · 1.7 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | assets/Vocos.pt | Weights | 54.4 MB | 09a670eda1c0 | | model-00001-of-00004.safetensors | Weights | 4.9 GB | de39e68ffea5 | | model-00002-of-00004.safetensors | Weights | 4.9 GB | 095633536dea | | model-00003-of-00004.safetensors | Weights | 4.3 GB | 4ea83267712f | | model-00004-of-00004.safetensors | Weights | 3.2 GB | f988c1730525 | | added_tokens.json | Configuration | 1.4 KB | — | | assets/chattts_tokenizer/special_tokens_map.json | Configuration | 7.8 KB | — | | config.json | Configuration | 3.4 KB | — | | configuration_minicpm.py | Configuration | 7.6 KB | — | | image_processing_minicpmv.py | Configuration | 16.7 KB | — | | model.safetensors.index.json | Configuration | 133.3 KB | — | | modeling_minicpmo.py | Configuration | 140.6 KB | — | | modeling_navit_siglip.py | Configuration | 42.1 KB | — | | preprocessor_config.json | Configuration | 714 B | — | | processing_minicpmo.py | Configuration | 20.0 KB | — | | resampler.py | Configuration | 35.6 KB | — | | special_tokens_map.json | Configuration | 5.4 KB | — | | tokenization_minicpmo_fast.py | Configuration | 3.0 KB | — | | utils.py | Configuration | 7.2 KB | — | | README.md | Documentation | 50.3 KB | — | | assets/Skiing.mp4 | Other | 8.5 MB | 479ace116d6a | | assets/demo.wav | Other | 1.5 MB | d0b347d8ed0b | | assets/input_examples/Trump_WEF_2018_10s.mp3 | Other | 161.1 KB | — | | assets/input_examples/assistant_default_female_voice.wav | Other | 224.0 KB | 2ee6f84892e6 | | assets/input_examples/assistant_female_voice.wav | Other | 235.2 KB | 1d712ba6de1d | | assets/input_examples/assistant_male_voice.wav | Other | 144.0 KB | e6b5eff26be1 | | assets/input_examples/audio_understanding.mp3 | Other | 321.0 KB | — | | assets/input_examples/chi-english-1.wav | Other | 492.2 KB | — | | assets/input_examples/cxk_original.wav | Other | 384.0 KB | — | | assets/input_examples/exciting-emotion.wav | Other | 696.0 KB | — | | assets/input_examples/fast-pace.wav | Other | 986.4 KB | — | | assets/input_examples/icl_20.wav | Other | 618.5 KB | 53892ece0713 | | assets/input_examples/indian-accent.wav | Other | 1.4 MB | 716533af9ec8 | | .gitattributes | Repository | 1.7 KB | — | | assets/chattts_tokenizer/tokenizer.json | Tokenizer | 448.6 KB | — | | assets/chattts_tokenizer/tokenizer_config.json | Tokenizer | 11.0 KB | — | | merges.txt | Tokenizer | 1.7 MB | — | | tokenizer.json | Tokenizer | 7.0 MB | — | | tokenizer_config.json | Tokenizer | 14.1 KB | — | | vocab.json | Tokenizer | 2.8 MB | — | ## License and Download License apache-2.0 Access Open weights, no gate Download size 17.4 GB [Download from OpenBMB](https://huggingface.co/openbmb/MiniCPM-o-2_6) Released by OpenBMB through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0). ## Built From - Described by arXiv:2405.17220 - Described by [arXiv:2408.01800](https://savrn.com/papers/minicpm-v-a-gpt-4v-level-mllm-on-your-phone) - Trained on (disclosed) openbmb/RLAIF-V-Dataset ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 17.4 GB | | 16-bit | 17.3 GB | | 8-bit | 8.7 GB | | 4-bit | 4.3 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Compare MiniCPM-o-2_6 - [gemma-4-E4B-it vs MiniCPM-o-2_6](https://savrn.com/models/compare/gemma-4-e4b-it-vs-minicpm-o-2-6) ## Questions About MiniCPM-o-2_6 ### How much GPU memory does MiniCPM-o-2_6 need? About 20.8 GB at 16-bit and 5.2 GB at 4-bit: the weights (8.7B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run MiniCPM-o-2_6 on? At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### Can I use MiniCPM-o-2_6 commercially? Yes. MiniCPM-o-2_6 is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors. ### What is MiniCPM-o-2_6's context length? 32,768 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Any to any ### [gemma-4-E4B-it-qat-w4a16-ct](https://savrn.com/models/gemma-4-e4b-it-qat-w4a16-ct) [Google](https://savrn.com/model-publishers/google) Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages. Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in five distinct sizes: E2B, E4B, 12B, 26B A4B, and 31B. Their diverse sizes make them deployable in environments ranging from… Open weights apache-2.0 8.7B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it-qat-w4a16-ct) Model · Any to any ### [gemma-4-E4B-it](https://savrn.com/models/gemma-4-e4b-it) [Google](https://savrn.com/model-publishers/google) Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on E2B, E4B, and 12B) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages. Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in five distinct sizes: E2B, E4B, 12B, 26B A4B, and 31B. Their diverse sizes make them deployable in environments ranging from… Open weights apache-2.0 8B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it) Model · Any to any ### [gemma-4-E4B-it-MLX-4bit](https://savrn.com/models/gemma-4-e4b-it-mlx-4bit) [LM Studio Community](https://savrn.com/model-publishers/lmstudio-community) LM Studio Community models highlights program. Highlighting new & noteworthy models by the community. Join the conversation on Discord. 4-bit quantized version of gemma-4-E4B-it using MLX, optimized for Apple Silicon. Special thanks to the Apple Machine Learning Research team for creating MLX. LM Studio is not the creator, originator, or owner of any Model featured in the Community Model Program. Each Community Model is created and provided by third parties. LM Studio does not endorse, support, represent or guarantee the completeness, truthfulness, accuracy, or reliability of any Community Model. You understand that Community Models can produce content that might be offensive, harmful… Open weights apache-2.0 8B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it-mlx-4bit) Model · Any to any ### [gemma-4-E4B-it-MLX-8bit](https://savrn.com/models/gemma-4-e4b-it-mlx-8bit) [LM Studio Community](https://savrn.com/model-publishers/lmstudio-community) LM Studio Community models highlights program. Highlighting new & noteworthy models by the community. Join the conversation on Discord. 8-bit quantized version of gemma-4-E4B-it using MLX, optimized for Apple Silicon. Special thanks to the Apple Machine Learning Research team for creating MLX. LM Studio is not the creator, originator, or owner of any Model featured in the Community Model Program. Each Community Model is created and provided by third parties. LM Studio does not endorse, support, represent or guarantee the completeness, truthfulness, accuracy, or reliability of any Community Model. You understand that Community Models can produce content that might be offensive, harmful… Open weights apache-2.0 8B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it-mlx-8bit) Model · Any to any ### [gemma-4-E4B-it-MLX-6bit](https://savrn.com/models/gemma-4-e4b-it-mlx-6bit) [LM Studio Community](https://savrn.com/model-publishers/lmstudio-community) LM Studio Community models highlights program. Highlighting new & noteworthy models by the community. Join the conversation on Discord. 6-bit quantized version of gemma-4-E4B-it using MLX, optimized for Apple Silicon. Special thanks to the Apple Machine Learning Research team for creating MLX. LM Studio is not the creator, originator, or owner of any Model featured in the Community Model Program. Each Community Model is created and provided by third parties. LM Studio does not endorse, support, represent or guarantee the completeness, truthfulness, accuracy, or reliability of any Community Model. You understand that Community Models can produce content that might be offensive, harmful… Open weights apache-2.0 8B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it-mlx-6bit) Model · Any to any ### [gemma-4-E4B-it-MLX-5bit](https://savrn.com/models/gemma-4-e4b-it-mlx-5bit) [LM Studio Community](https://savrn.com/model-publishers/lmstudio-community) LM Studio Community models highlights program. Highlighting new & noteworthy models by the community. Join the conversation on Discord. 5-bit quantized version of gemma-4-E4B-it using MLX, optimized for Apple Silicon. Special thanks to the Apple Machine Learning Research team for creating MLX. LM Studio is not the creator, originator, or owner of any Model featured in the Community Model Program. Each Community Model is created and provided by third parties. LM Studio does not endorse, support, represent or guarantee the completeness, truthfulness, accuracy, or reliability of any Community Model. You understand that Community Models can produce content that might be offensive, harmful… Open weights apache-2.0 8B parameters 131,072 tokens transformers [View model](https://savrn.com/models/gemma-4-e4b-it-mlx-5bit) ## OpenBMB [All models and datasets](https://savrn.com/model-publishers/openbmb) ## Versions - [06849bfd36da](https://savrn.com/models/minicpm-o-2-6/versions/06849bfd36da) · current 2026-09-18 ## Explore More - [All any to any models](https://savrn.com/models/tasks/any-to-any) - [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-20. - [Hugging Face record](https://huggingface.co/openbmb/MiniCPM-o-2_6) - [How the hub is built](https://savrn.com/model-hub/methodology)