MiniCPM5-1B

Item: MiniCPM5-1B
Rating: 2
Author: Ali Madjaji

OpenBMB

Per the published Apache 2.0 LICENSE, the 1.08B-parameter weights carry no use restrictions and are well suited to on-device or edge deployments. Unusually for a Chinese-origin release, OpenBMB ships a full data card naming every pre-training and SFT corpus, which closes the AI Act Art. 53 transparency gap most peers leave open. The remaining hedge for regulated EU buyers is vendor jurisdiction (Tsinghua-affiliated lab, mainland China) — self-host the weights and treat any vendor-side service as out of scope.

Sovereignty

Licence: Apache 2.0Commercial: UnrestrictedTraining data: DocumentedOrigin: China

Licence facts

Parameters: 1.08B (679M non-embedding)
Architecture: Dense Llama-compatible decoder, 24 layers, GQA (16 Q / 2 KV)
Context window: 131,072 tokens
Languages: English, Chinese
Modality: Text-only
Reasoning mode: Hybrid thinking / non-thinking toggle via chat template

Performance & pricing?

Known risks

Tsinghua-affiliated Chinese vendor: EU deployers should self-host and treat vendor-side services as out of scope for personal-data workloads.
Cross-border transfer of any prompt or fine-tuning data to vendor APIs would trigger Chapter V GDPR scrutiny — keep the model on EU infrastructure.
Model card explicitly warns of potential inaccurate, biased or unsafe outputs; deployers must layer their own safety filters for high-stakes use.

Sources

→ Model Card → Ultra-FineWeb dataset → UltraData-SFT-2605 dataset

Compare with

vsStep-3.7-Flash vsHy-MT2-30B-A3B vsMixtral 8x22B Instruct vsMistral Nemo

Reviewed by Ali Madjaji · Last reviewed 2026-05-30· Reviewed 23 days agoSuggest a correction