MiniCPM5-1B brings 128K on-device context to phones

OpenBMB released MiniCPM5-1B, a one-billion-parameter language model built to run on consumer phones. The model supports native tool calling and the Model Context Protocol (MCP) and carries a 128K-token context window. MiniCPM5-1B is available on Hugging Face under an Apache 2.0 license and is compatible with vLLM, SGLang and standard Transformers inference libraries.

The 128K-token context equals roughly 96,000 words of continuous text, which OpenBMB says enables long roleplay sessions, full-document digests and agent contexts that persist across many exchanges. The release is aimed at local agent workflows where a phone or laptop can call local tools, query a calendar, search a local database or reach an MCP research server without sending user data to a cloud API.

The model builds on the MiniCPM4 architecture and introduces InfLLM v2, a trainable attention mechanism that restricts each token’s attention to fewer than 5% of surrounding tokens during long-context inference. OpenBMB reported that this design reduces computation for long inputs while maintaining accuracy. Training used a filtering pipeline named UltraClean and roughly 8 trillion training tokens.

OpenBMB applied post-training optimization that combined reinforcement learning with efficient distillation from larger models. The company reported a 16-point increase on math, code and instruction-following benchmarks and a 29-percentage-point reduction in runaway-length responses after these steps.

In internal and independent evaluations, MiniCPM5-1B scored an average of 42.57 on a capability benchmark that covers general knowledge, domain knowledge, coding, instruction-following, math reasoning, logical reasoning and agentic tasks. The next-best 1B-class competitor scored 35.61 on the same benchmark. OpenBMB lists competitors at this scale including Alibaba’s Qwen3-0.6B and Qwen3.5-0.8B and Liquid AI’s LFM2.5-1.2B-Thinking.

Tests validated the model’s native tool calling and MCP support. When connected to a research MCP server, the model fetched a live Bitcoin price and returned three stock suggestions-Amazon, Microsoft and Nvidia-via the tool chain.

Limitations appeared in quick evaluations. In a classic logic-trap prompt about a man marrying his widow’s sister, the model produced a jurisdictional analysis and did not flag the paradox that a man with a widow is deceased. In a separate A/B decisional test the model hedged rather than choosing a side, a behavior noted for smaller conversational models.

MiniCPM5-1B is smaller than many modern models: Google’s Gemma 4 starts at an effective 2 billion parameters and scales to 31 billion, and Llama 4 Scout runs with 17 billion active parameters. OpenBMB describes MiniCPM5-1B as built for on-device agent use where long context and local tool integration are prioritized over larger model scale.