MiniCPM

MiniCPM is an open-source language model series designed for on-device deployment and resource-constrained scenarios.

MiniCPM provides a series of compact dense models, sparse-and-linear hybrid attention models, and quantized variants built for local deployment and resource-constrained scenarios. The models range from 0.5B to 8B parameters, including the MiniCPM5-2B and MiniCPM5-1B releases. They are built for local assistants, coding agents, tool-use workflows, and reasoning tasks. The system utilizes training datasets like UltraData, UltraX, UltraData-Code, UltraData-SFT-Agent-2609, and UltraData-RL-2609, along with post-training methods such as SFT, RL, and On-Policy Distillation. The models support long-context understanding, mathematics, coding, and tool use with native context lengths up to 131,072.

MiniCPM at a glance

Best for: Software engineer, Researcher
Runs on: self-hosted, open source
Pricing: Free
Released: January 2024
Category: AI chat assistant
Website: github.com/OpenBMB/MiniCPM

Key features

Dense 1B and 2B parameter model sizes
Long-context understanding up to 131,072 context length
On-device and resource-constrained deployment
Code and math reasoning capabilities
Tool use and agentic workflows
Sparse-and-linear hybrid attention modeling
Quantized formats including GGUF, MLX, GPTQ, and AutoAWQ

Use it for

Running local AI assistants on edge devices
Executing coding tasks and code reasoning
Solving math reasoning problems
Building tool-use and agentic workflows
Handling long-context document understanding

Alternatives to MiniCPM

AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
LMDeploy — Accelerates inference with high-performance engines and quantization tools.
SmallCode — Runs fully local with budget-managed context and a forgiving parser for small models.
claude-mem — It shares persistent, searchable agent memory across sessions and devices.
SimpleMem — Provides semantic lossless compression to reduce token consumption and maintain long-term context.
Every number in a tiny GPT — Inspects every lookup, matrix multiplication, and normalisation in a live transformer.
LM Studio — Runs open source language models locally with real-time voice transcription.
mini-SWE-agent — It offers high performance with a simplified, minimal codebase.
XTuner — High-efficiency training for 200B+ scale MoE models without expert parallelism.
LibreChat — Unifies multiple AI models and tools in one customizable, open-source platform.
llmspy.org — Self-host access to 530+ AI models across 24 providers with built-in tools.
Mem0 — It provides drop-in persistent memory for AI agents without changing pipelines.