Can you own it?
This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Intended & out-of-scope use
The publisher positions the model for "local assistants, coding agents, tool-use workflows, and reasoning scenarios where a compact model is preferred" and for "on-device, local deployment, and resource-constrained scenarios" (ev-card). The GitHub repository ships a desktop-pet demo that runs it locally (ev-github). The model is text-only. The card says generated content "should be reviewed and verified before use in high-stakes settings", so high-stakes decisions are out of scope without the deployer's own controls. Only English and Chinese are documented.
Known limitations, bias & failure modes
- No safety evaluation exists. The card's only safety text says the model "may produce inaccurate, biased, or unsafe outputs" and leaves safeguards to the user (
ev-card). - Benchmark rows are mixed. On the publisher's own table the model is behind LFM2.5-1.2B on GPQA-Diamond (26.26 vs 34.85), IFEval (80.41 vs 84.84), Multi-IF (43.54 vs 55.61) and MultiChallenge (19.48 vs 23.28). It is behind both Qwen baselines on BFCLv4 (25.15 vs 25.43 and 25.53) (
ev-bench). - Capacity is small. A 1.08B model scores 48.85 on MMLU-Pro in that table, so expect weak factual recall and long-tail knowledge.
- Only two languages are documented. The card lists English and Chinese, and nothing else is documented or evaluated.
- The decay-phase pretraining tier was synthesised with MiniCPM4 and Qwen3 (
ev-datasets), so the model carries those generators' errors and biases.
Openness tier & components
open_weights_recipe, as a judgement call between the 3 and 4 anchors. The open parts are Apache-2.0 weights at four stages (Base, SFT-only, final, plus official GGUF and MLX), a recipe description on the card, and three pretraining datasets. The card names the three datasets as the training corpus, and they are public and Apache-2.0 tagged on the Hub: Ultra-FineWeb, Ultra-FineWeb-L3 and UltraData-Math (ev-card, ev-datasets). That is more training data than most open-weight models release.
The closed parts are these. The SFT dataset UltraData-SFT-2605 sits behind an automatic Hub access gate and its card is unavailable (ev-sft-dataset). The synthesised RLVR data and the pairwise RLHF data are not linked. The repository holds user-side fine-tuning and quantisation scripts but no pretraining, RL or OPD pipeline and no evaluation harness (ev-github). There is no MiniCPM5 technical report, because the card's "Tech Report" link is the MiniCPM4 paper (ev-papers).
Components: weights open, licence open, training data partial, training code partial, evaluation partial, documentation partial.
License terms & permitted use
Apache License 2.0 for all MiniCPM5 checkpoints. The GitHub LICENSE file matches the canonical apache.org text line for line; only the copyright line differs ("Copyright [OpenBMB]"), so there is no field-of-use, scale, branding or registration clause (ev-license). The Hub metadata of the 1B and 2B families all reads apache-2.0 and ungated (ev-license-meta). Commercial use, modification, redistribution and sublicensing are permitted under the standard Apache terms (attribution, notice, patent grant and termination).
Two points matter. First, the LICENSE file is in the GitHub code repository. The Hub model repository has no LICENSE file and relies on the card sentence and metadata, so keep a copy of the GitHub file with any redistribution. Second, earlier MiniCPM generations (the 2B and V-2 era) have their weights under a custom General Model License that requires commercial authorisation by registration (ev-legacy-gml). The MiniCPM5 README replaces that with Apache-2.0. An older checkpoint carries different terms, so check before assuming this dossier applies.
The datasets are also Apache-2.0 tagged, but the Ultra-FineWeb card warns that its source datasets have their own licences (ev-datasets).
Supply-chain & provenance
The canonical source is huggingface.co/openbmb/MiniCPM5-1B, linked from the OpenBMB/MiniCPM GitHub README. The openbmb Hub org is not verified (isVerified: false), so identity rests on the GitHub cross-links and on Artificial Analysis attributing the model to OpenBMB (ev-repo-files, ev-aa).
Weights are a single safetensors shard with an LFS sha256, and there is no pickle file. The Hub's own scan reports no flagged files. There is no signature or attestation. Official GGUF (F16, Q8_0, Q4_K_M) and MLX 4-bit conversions come from the same org. Many community conversions exist, including repos named "Uncensored", which are uninspected.
On data lineage, Ultra-FineWeb-L3 was generated with MiniCPM4 and Qwen3, and the math selection used unnamed proprietary models (ev-datasets). The checkpoint checklist scores 5 of 8.
EU AI Act posture
The model is a general-purpose AI model. At 1.08B parameters it would need roughly 1.5 quadrillion training tokens to reach 1e25 FLOPs (6 x parameters x tokens), so it is not a systemic-risk model and no Article 55 package applies. Under Apache-2.0 with public weights it plausibly qualifies for the Article 53 open-source exemption from the Annex XI and XII documentation duties.
The surviving duties are a copyright policy and a public training-content summary. OpenBMB states no copyright policy and no EU-specific statement. The dataset cards are public and describe the pretraining sources, which partly serves as a training-content summary. The SFT card is gated, no pretraining token count is given, and the source-dataset licence caveat applies. Whether OpenBMB or Modelbest has signed the GPAI Code of Practice is unknown.
A deployer who fine-tunes and places the result on the EU market may become the provider for that derivative and should document the gaps above.
Benchmarks & evaluation
All numbers come from one publisher image on the model card (ev-bench). It covers Thinking mode only, against three small thinking-mode models (Qwen3-0.6B, Qwen3.5-0.8B, LFM2.5-1.2B), with no harness or prompt detail and no non-thinking results. The reported figures are an average of 42.57 against 35.61 for LFM2.5-1.2B, MATH-500 91.60, AIME-2025 and AIME-2026 both 40.42, LCB-v6 33.52, MMLU-Pro 48.85, tau2-Bench Telecom 79.53 and BFCLv4 25.15.
The card says the advantage is most visible in tool use, but on BFCLv4 the model is level with or below the Qwen baselines, so the tool-use lead rests on one benchmark. The comparison set has no Gemma, Llama, SmolLM, Granite or Qwen3-1.7B baseline. Artificial Analysis lists the model but publishes no score (ev-aa), and no other independent evaluation exists. The table is an unverified publisher claim, so run an evaluation on the deployer's own task.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?
StrongApache-2.0 across Base, SFT, final, GGUF and MLX checkpoints, ungated, with commercial use allowed and no field-of-use or registration clause.
TransparencyDo you know what it is: weights, training, behaviour, and legible terms?
ModerateThe recipe and three pretraining datasets are public, but the SFT dataset card is gated and unread, the RL data is partly unreleased, there is no MiniCPM5 report or training pipeline, and the benchmark table is an unverified image.
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.