MiniCPM4
Ultra-efficient 8B LLM built for on-device inference
About
MiniCPM4 is OpenBMB's ultra-efficient 8B language model built for on-device use, released June 2025 under Apache 2.0. It was trained on 8 trillion tokens with a 32K native context extendable to 128K via YaRN, and the team reports roughly 7x faster decoding than Qwen3-8B on a Jetson AGX Orin. A 0.5B sibling covers the smallest devices; both trade some peak capability for speed and low memory use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: ~7x decoding speedup vs Qwen3-8B on Jetson AGX Orin; strong 128K needle-in-a-haystack retrieval.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Released 2025-06-06 by OpenBMB under Apache-2.0 (confirmed on HF model card); 8B flagship trained on 8T tokens; 32K native context, 128K via YaRN; 0.5B sibling also released; Speed claims (7x vs Qwen3-8B on Jetson AGX Orin) are team-reported from the HF card and labeled as such. License: Apache-2.0 confirmed on the Hugging Face card..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Released 2025-06-06 by OpenBMB under Apache-2.0 (confirmed on HF model card)
- · 8B flagship trained on 8T tokens; 32K native context, 128K via YaRN; 0.5B sibling also released
- · Speed claims (7x vs Qwen3-8B on Jetson AGX Orin) are team-reported from the HF card and labeled as such