MiniCPM4
OpenBMB's edge-first LLM family: an 8B flagship and a 0.5B tiny model with 5x generation speedups on device chips.
About
MiniCPM4 is OpenBMB's 2025 model family built explicitly for end-side devices, released June 6, 2025. The flagship MiniCPM4-8B trains on 8 trillion tokens and the small MiniCPM4-0.5B on 1 trillion, both under Apache 2.0. The efficiency story has four parts: InfLLM v2 trainable sparse attention (each token attends to under 5% of tokens in long contexts), FP8 training with multi-token prediction, the UltraClean/UltraChat data pipelines, and the CPM.cu and ArkInfer inference systems.
OpenBMB claims over 5x generation acceleration on typical edge chips with optimal performance at the same scale, and the model natively handles 32,768 tokens, extendable to 131,072 with the validated LongRoPE factors. The family is heavily packaged for deployment: GGUF, MLX, QAT Int4, Eagle speculative heads, and MCP/survey variants all ship in the Hugging Face collection. The technical report (arXiv:2506.07900) documents the design, though headline efficiency numbers are provider-measured.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No canonical independent benchmark table found; OpenBMB claims optimal performance at the same scale with 5x+ generation acceleration on edge chips (provider-reported, technical report arXiv:2506.07900)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: real OpenBMB release Jun 6, 2025 with Apache-2.0 weights, published technical report (arXiv:2506.07900), and unusually complete deployment packaging (GGUF/MLX/QAT/Eagle heads); followed by MiniCPM4.1 and MiniCPM5 showing active maintenance. Open weights. Strengths: serious edge-inference engineering, permissive license, healthy download counts (11K on the 0.5B). Docked for: (1) efficiency and quality claims are provider-measured with no independent replication found; (2) sparse-attention and speculative stacks add integration complexity versus dense models.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · OpenBMB release 2025-06-06; 8B (8T tokens) + 0.5B (1T tokens)
- · Technical report arXiv:2506.07900; follow-ups MiniCPM4.1, MiniCPM5, MiniCPM-SALA
- · Apache-2.0 per model card; GGUF/MLX/QAT/Eagle deployment packaging
- · Efficiency claims are provider-measured; sparse-attention stack adds integration complexity