OpenBMB/MiniCPM

MiniCPM5-1B: A SOTA 1B on-device LLM, small yet powerful.

What it solves

MiniCPM5-1B is a compact, 1B-parameter dense Transformer model designed for on-device and local deployment in resource-constrained scenarios. It aims to provide state-of-the-art (SOTA) performance for its size class, particularly in reasoning, coding, and tool-use workflows where a small memory footprint is required.

How it works

Built on the standard LlamaForCausalLM architecture, the model is trained using a three-stage process: base training, mid-training, and post-training. The post-training phase utilizes a combination of Supervised Fine-Tuning (SFT) with deep-thinking and hybrid-thinking data, Reinforcement Learning (RL) to improve reasoning and reliability, and On-Policy Distillation (OPD) to distill knowledge from specialized RL teachers into a single release model.

It features a hybrid reasoning system where a single checkpoint can act as either a fast assistant or a deliberate reasoner via a built-in <think> chat template and an enable_thinking toggle.

Who it’s for

  • Local AI developers: Those building on-device assistants, coding agents, or tool-use workflows.
  • Edge computing users: Users with limited hardware resources who need a high-performance LLM.
  • ML engineers: Developers looking for a model that integrates easily with mainstream inference engines like vLLM, SGLang, and Transformers without custom kernels.

Highlights

  • 1B-class SOTA: Outperforms other open-source models of similar size in reasoning, code, and agentic tool use.
  • Hybrid Reasoning: Supports both "Think" and "No Think" modes within the same model checkpoint.
  • Standard Architecture: Uses LlamaForCausalLM, ensuring compatibility with most major inference backends (vLLM, SGLang, llama.cpp, etc.).
  • Agent-Ready Deployment: Provides specialized "Agent Skills" for Cursor and Claude Code to automate deployment and fine-tuning workflows.
  • Native Tool Calling: Emits XML-style tool calls, with native support in SGLang for OpenAI-compatible conversion.

Related

  • Project
    OpenBMB/MiniCPM-Robot

    A family of efficient on-device AI models for robots, featuring a 1.5B VLA for generalist manipulation and a 0.9B model for real-time embodied target tracking.

  • Dispatch
    Llama 3.2 Release Notes: Multimodal Vision and On-Device Small Language Models

    Meta has released Llama 3.2, introducing multimodal vision capabilities in 11B and 90B sizes and lightweight 1B and 3B text-only models optimized for on-device deployment.

  • Project
    jmaczan/tiny-vllm

    tiny‑vllm is a C++/CUDA LLM inference engine (a smaller, educational version of vLLM) that loads a Llama 3.2 1B Instruct model from safetensors, runs the full forward pass on GPU (prefill + decode), implements KV‑cache, static/continuous batching, FlashAttention‑style online softmax and PagedAttention, and includes a step‑by‑step tutorial for learning or teaching GPU‑based LLM serving.

  • Project
    OpenBMB/MiniCPM-V-Apps

    A collection of on-device mobile demos for the MiniCPM-V family of multimodal models, enabling local AI inference on iOS, Android, and HarmonyOS.

  • Project
    zai-org/GLM-5

    GLM‑5 is an open‑source series of very large LLMs (320 B‑744 B total parameters) optimized for multimodal input, 1 M‑token contexts, and long‑horizon agentic tasks. It introduces hybrid sparse/linear attention, IndexShare, and a custom RL engine (slime) to achieve frontier‑level performance on coding and planning benchmarks while offering cheap FP8 inference. Models are downloadable from Hugging Face/ModelScope and can be served with SGLang, vLLM, KTransformers, etc., with optional ‘reasoning_effort’ control.