MiniCPM5

MiniCPM5 1B

MiniCPM5-1B is the first release in the MiniCPM5 series, a dense 1B-parameter causal language model built for on-device, local deployment, and resource-constrained scenarios. It is based on the standard LlamaForCausalLM architecture and supports English and Chinese. According to its official model card, it reaches 1B-class open-source SOTA within its comparison set, with advantages noted in agentic tool use, code generation, and difficult reasoning. It provides native long-context support and a hybrid reasoning mode via a built-in `` chat template that can be switched with `enable_thinking`. The checkpoint is licensed under Apache-2.0 and is available alongside related SFT, Base, GGUF, and MLX variants.

Model size

1.08B

Format

FP16 weights

Context

131,072

MiniCPM5 1B Base

MiniCPM5 1B Base is the pre-training-only base checkpoint in the MiniCPM5 series, released under the apache-2.0 license. It is a dense causal language model using a standard LlamaForCausalLM architecture, with 1,080,632,832 total parameters and 679,552,512 non-embedding parameters. It has 24 layers, GQA with 16 query heads and 2 KV heads, and a context length of 131,072. The model targets English and Chinese, supports text-generation, and is associated with long-context, tool-calling, on-device, and edge-ai tags. Pretraining data includes openbmb/Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math, and UltraData-SFT-2605. The official model card also notes a built-in think chat template switchable via enable_thinking for the released MiniCPM5-1B model, but this base checkpoint is pre-training only.

Training checkpointResearch/fine-tuning checkpoint, not the same final chat release.

Model size

1.08B

Format

FP16 weights

Context

131,072

MiniCPM5 1B SFT

MiniCPM5 1B SFT is an Apache-2.0 licensed, text-generation model from openbmb/MiniCPM5-1B-SFT. It is an SFT-only checkpoint (before RL/OPD) in the MiniCPM5 series, supporting English and Chinese. The model is a dense 1B Transformer with a LlamaForCausalLM architecture, 1,080,632,832 parameters (679,552,512 non-embedding), 24 layers, GQA with 16 Q heads and 2 KV heads, and a 131,072 token context length. It includes a hybrid reasoning chat template with a `thinking` mode activated via `enable_thinking`, and is intended for on-device, local deployment, and resource-constrained scenarios.

Training checkpointResearch/fine-tuning checkpoint, not the same final chat release.

Model size

1.08B

Format

FP16 weights

Context

131,072

MiniCPM5 2B GGUF

MiniCPM5 2B GGUF is a GGUF-formatted text-generation model from OpenBMB, licensed under Apache-2.0. It is described as a dense 2B Transformer, the second model in the MiniCPM5 series after MiniCPM5-1B, and is intended for on-device, local deployment, and resource-constrained scenarios. The model card lists English and Chinese language support and tags for long-context, tool-calling, on-device, and edge-ai use. Training data references include the Ultra-FineWeb, UltraX-Preview, Ultra-FineWeb-L3, UltraData-Math, UltraData-Code, UltraData-SFT-2605, UltraData-SFT-Agent-2609, and UltraData-RL-2609 datasets.

Format

Q4_K_M GGUF weights

MiniCPM5 2B

MiniCPM5-2B is an Apache-2.0 licensed dense 2B Transformer text-generation model from OpenBMB, released as the second model in the MiniCPM5 series after MiniCPM5-1B. The model card describes it as designed for on-device, local deployment, and resource-constrained scenarios, supporting English and Chinese, with tags including long-context and tool-calling. It is intended to be used with transformers and is distributed with safetensors weights. Training data references listed datasets such as Ultra-FineWeb, UltraX-Preview, UltraData-Math, UltraData-Code, and UltraData SFT/Agent/RL collections. The card claims 2B-class open-source SOTA within its comparison set and competitiveness with 4B-class models, with stated advantages in coding, mathematics, long-context understanding, tool use, and agentic tasks.

Model size

2.52B

Format

FP16 weights

Context

131,072

MiniCPM5 2B Base

MiniCPM5 2B Base is a dense 2B Transformer text-generation model from the MiniCPM5 series, released under the Apache-2.0 license by openbmb. It is positioned for on-device, local deployment, and resource-constrained scenarios, and its model card describes it as reaching 2B-class open-source SOTA within its stated comparison set. The model supports English and Chinese, uses the transformers library, and is associated with long-context, tool-calling, on-device, and edge-ai tags. It was trained on datasets including Ultra-FineWeb, UltraX-Preview, UltraData-Math, UltraData-Code, and several UltraData SFT/RL datasets.

Training checkpointResearch/fine-tuning checkpoint, not the same final chat release.

Model size

2.52B

Format

FP16 weights

Context

524,288

MiniCPM5 2B GPTQ

MiniCPM5 2B GPTQ is a GPTQ-quantized text-generation model from OpenBMB, released under the Apache 2.0 license. It is described as a dense 2B Transformer intended for on-device, local deployment, and resource-constrained scenarios, and is the second model in the MiniCPM5 series following MiniCPM5-1B. The official model card states it reaches 2B-class open-source SOTA within its comparison set and remains competitive with 4B-class models overall, with reported advantages over comparable-size models in coding, mathematics, long-context understanding, tool use, and agentic tasks. Supported languages listed in the card are English and Chinese. Training data listed includes Ultra-FineWeb, UltraX-Preview, Ultra-FineWeb-L3, UltraData-Math, UltraData-Code, UltraData-SFT-2605, UltraData-SFT-Agent-2609, and UltraData-RL-2609.

Model size

2.52B

Format

Q4 weights

Context

131,072

MiniCPM5 2B Midtrain

MiniCPM5 2B Midtrain is the second model in the MiniCPM5 series, a dense 2B Transformer from OpenBMB built for on-device, local deployment, and resource-constrained scenarios. It scales up the same training recipe used for MiniCPM5-1B and is released under the apache-2.0 license. The model card describes it as reaching 2B-class open-source SOTA within its comparison set and remaining competitive with 4B-class models overall, with reported advantages over comparable-size models in coding, mathematics, long-context understanding, tool use, and agentic tasks. Languages listed are English and Chinese, and the model is associated with the transformers library and text-generation pipeline. Training data includes openbmb Ultra-FineWeb, UltraX-Preview, Ultra-FineWeb-L3, UltraData-Math, UltraData-Code, UltraData-SFT-2605, UltraData-SFT-Agent-2609, and UltraData-RL-2609.

Training checkpointResearch/fine-tuning checkpoint, not the same final chat release.

Model size

2.52B

Format

FP16 weights

Context

131,072

MiniCPM5 2B MLX

MiniCPM5 2B MLX is an Apache-2.0 licensed, text-generation model from OpenBMB, designed as a dense 2B Transformer for on-device, local deployment, and resource-constrained scenarios. According to the model card, it is the second model in the MiniCPM5 series and supports English and Chinese. Its stated focus areas include coding, mathematics, long-context understanding, tool use, and agentic tasks, and it is presented as achieving SOTA performance within a comparison set of similarly sized open-source models while remaining competitive with 4B-class models overall.

Model size

2.52B

Format

Q4 weights

Context

131,072

MiniCPM5 2B SFT

MiniCPM5 2B SFT is a dense 2B Transformer text-generation model in the MiniCPM5 series, released by OpenBMB under the apache-2.0 license. According to its official model card, it follows the training recipe of MiniCPM5-1B, supports English and Chinese, and is tagged for long-context, tool-calling, on-device, and edge-AI use. It uses the transformers library and safetensors format, with pipeline tag text-generation. The model card states it targets on-device, local deployment, and resource-constrained scenarios, and describes it as achieving SOTA performance among compared 2B-class open-source models while remaining competitive with 4B-class models in certain evaluated areas. Training data includes Ultra-FineWeb, UltraX-Preview, Ultra-FineWeb-L3, UltraData-Math, UltraData-Code, and several UltraData SFT/RL datasets. No specific context length, parameter count beyond '2B', hardware requirements, or benchmarking details are provided in the source facts.

Training checkpointResearch/fine-tuning checkpoint, not the same final chat release.

Model size

2.52B

Format

FP16 weights

Context

131,072