MiniCPM5-2B for sidekick (Core ML)
openbmb/MiniCPM5-2B, a Llama (MiniCPM5) chat model, converted to Core ML by sidekick for Apple silicon's Neural Engine. This repository holds one build, palette8/, graded against the original model on every compute path it runs on.
Licence: Apache-2.0 · publishing tier 2a. Apache-2.0 weights, trained on openbmb's UltraData sets, which are Apache-2.0 as collections but defer to their sources' terms: web and code corpora whose terms weren't checked, synthetic data from unnamed generators, and RL problems under CC BY-SA 4.0. Share-alike, unverified and unnamed-generator data: 2a, with a notice on the card; no terms travel with the weights, so no gate. Whether a dataset's terms reach a model trained on it is legally unsettled; the training data below gives each source's licence. How sidekick decides what it publishes.
Download
| Build | Weights | Decoder | Context (tokens) | Download | Runs on (grade) |
|---|---|---|---|---|---|
palette8/ |
8-bit palettes (GPTQ) | 4 programs | 2033 | 2.83 GB | Neural Engine B |
sidekickd fetch minicpm5-2b
Needs sidekick 0.17.0 or later. How fetch works · Installing without fetch · Calling the model · Why compiled programs
Fidelity
Graded against the original model in fp32 by sidekick's converter and grader, on the KL divergence of each next-token distribution and top-1 agreement over held-out texts, per text family, the build earning its worst family's letter. What the grades mean.
B on the Neural Engine, Apple M1 Max, macOS 27.0 (both calibration draws screened alike):
| Texts | Positions | KL mean | KL p99 | Top-1 | Grade |
|---|---|---|---|---|---|
| Long prose (held out) | 470 | 1.28e-03 | 1.17e-02 | 99.57% | B |
| Repository texts | 1032 | 2.89e-04 | 2.07e-03 | 98.93% | B |
Task check, the build served by sidekickd against the original model in fp32 on the same requests of sidekick's chat benchmark: passed.
| Task | Requests | This build | fp32 | Difference | Identical replies |
|---|---|---|---|---|---|
| Titles | 20 | 0.448 | 0.441 | +0.007 | 95% |
| Tags (fixed vocabulary) | 20 | 0.792 | 0.792 | +0.000 | 100% |
| Extraction | 24 | 0.910 | 0.896 | +0.014 | 92% |
| Instructions | 20 | 1.000 | 1.000 | +0.000 | 90% |
Graded on Apple M1 Max only; other chips haven't been graded.
The grading report's file hashes match every file of this build but chat.toml, and its digest of chat.toml matches this one's content without graded_os, the line that records where it was graded.
Before you install
- Neural Engine compile cache: the first load of an install compiles the decoder for the Neural Engine, and macOS keeps the compiled program in a system cache that you can't delete: 2.14 GiB, about one copy of its weights (2.12 GiB), measured on
palette8/'s first load (Apple M1 Max, macOS 27.0, 2026-10-29). It is paid once per install and again after a macOS update, and a sidekickd signed with another identity than the release's compiles it again. - Shared with Apple Foundation Models: the Neural Engine is shared, so other apps using Apple Foundation Models, or other Neural Engine work, stall this model's generations while they run. sidekickd reports the stalls but can't prevent them.
Open chat models on the Neural Engine
Training data
Training data: partly disclosed. Each dataset under its original source's licence:
| Dataset | Licence | Terms | Stage |
|---|---|---|---|
| Ultra-FineWeb | Apache-2.0 (the collection); its sources' own terms apply | Unknown | mid-training |
| UltraX-Preview | Apache-2.0 (the collection); its sources' own terms apply | Unknown | mid-training |
| Ultra-FineWeb-L3 | Apache-2.0 | Unknown | mid-training |
| UltraData-Math | Apache-2.0 | Unknown | mid-training |
| UltraData-Code | Apache-2.0 (the collection); each source repository's licence applies | Unknown | mid-training |
| UltraData-SFT-2605 | Apache-2.0 (the collection); upstream licences apply | Unknown | SFT |
| UltraData-SFT-Agent-2609 | Apache-2.0 (the collection); upstream licences apply | Unknown | SFT |
| UltraData-RL-2609 | Apache-2.0 (the collection); upstream MIT, CC BY 4.0 and CC BY-SA 4.0 apply | Share-alike | RL and on-policy distillation |
Not confirmed from a primary source
- The terms of the Chinese web sources behind Ultra-FineWeb (WuDao, SkyPile, MiChao, IndustryCorpus2) and of UltraX's sources were not checked.
- The card doesn't say which datasets went into base training and which into mid-training.
Original model
| Model | openbmb/MiniCPM5-2B |
| Revision | f97400052a43d642bbc6e9975e2397e3ae6a6b52 |
| Licence | Apache-2.0 |
| Licence file | LICENSE: the Apache License 2.0 as the Apache Software Foundation publishes it; the model's repository carries no licence file |
| Base model | openbmb/MiniCPM5-2B-Base (Apache-2.0) |
| Changes in this repository | Converted to Core ML in 8-bit palettes (gptq) by tools/convert_causal_lm.py, with a sidekick manifest added. Reproduce it |
Checksums
palette8/: ledger build minicpm5-2b/coreml/palette8/fffc316c8db0, 25 files
sidekickd fetch checks every file against sidekick's checksum ledger; palette8.sha256 has the same list for shasum -a 256 -c.
| Built with | |
|---|---|
| Converter | tools/convert_causal_lm.py at 9df79a88e3ad |
| coremltools | 9.0 |
| PyTorch | 2.13.0 |
| Python | 3.12.13 |
| OS | macOS 27.0 (26A428) |
| Xcode | 27.1 (27A9275) |
| Converted on | Apple M1 Max |
| Converted | 2026-10-28 |
| sha256 | Size | Path |
|---|---|---|
c9e8e343366783d8a6fe55a42661c0bc51b92b0308dfc77f672266de5ffe023b |
3.1 kB | chat.toml |
bc1f78ef8a86b6bdf64a056cf89dd0ac506db4f198ab363d85168e2500ab44dd |
5.4 kB | chat_template.jinja |
57e87c602f436ae26d5bac8d94ecdbc8332eab67468f4f292cfbcc5b4687c2ad |
243 B | decoder.0.mlmodelc/analytics/coremldata.bin |
b6fbd1598f9d88f2e5f290d7404f291aacb74f87b7d18a045560e1e864ff90f7 |
522 B | decoder.0.mlmodelc/coremldata.bin |
ac3cb0d6fcbe9bea48ecbe191bcc631c2d45be3db3725c2e40d742cc0d7b45d4 |
4.0 kB | decoder.0.mlmodelc/metadata.json |
71e2e6ac88fb2c4e15773ef4ed41fbf4148bf0c065adbe55de060164ecc736bb |
440.0 kB | decoder.0.mlmodelc/model.mil |
3cc005bc652a2f48a62439d947e3e432295331d1e3cd346fbd496e666930b103 |
669.1 MB | decoder.0.mlmodelc/weights/weight.bin |
fad814b42dd3b60d0d359d26c9b045571192008691373ce81815a1fb3851750a |
243 B | decoder.1.mlmodelc/analytics/coremldata.bin |
65609a6d6b4384f9d6a181f5955cab965f6e6f525d4907388c343a606a2820fc |
522 B | decoder.1.mlmodelc/coremldata.bin |
c9f849aef7adc84fb92d176b44f601f0d6d66b23d0c3bcb3b771d5466ab7f524 |
4.0 kB | decoder.1.mlmodelc/metadata.json |
1bcabca14e7fbaebfefd7a74d6db0d690c942caa6eb504eafc7206bd0097914e |
444.2 kB | decoder.1.mlmodelc/model.mil |
5f32348e9a7199a8a29df00dee050d8aff21a089a5a8bb5ff422f9d9fb7c5eb2 |
669.1 MB | decoder.1.mlmodelc/weights/weight.bin |
ee208fc01e14dbd29b9f3668586fc802b600ab8d2044dc24bf15f630fcb1479f |
243 B | decoder.2.mlmodelc/analytics/coremldata.bin |
cf52a77ccba10342874320a718a5b197d0fdad0fc63976e022476aa82021c720 |
522 B | decoder.2.mlmodelc/coremldata.bin |
eba0e61a63a11017b67459500fb92259176739c922aa015badc3c2a27fa5f124 |
4.0 kB | decoder.2.mlmodelc/metadata.json |
90829f1c16e3fb9b6720ce87c24f74f3bf8482e94717e3f60271dbac5528e1f0 |
316.6 kB | decoder.2.mlmodelc/model.mil |
d75635a1d98adc7a861bd062e2ecdeb0d15cddf7279b59dc2a859a6209b310fd |
477.9 MB | decoder.2.mlmodelc/weights/weight.bin |
0413cd4bf35fbfcfb6098837e5f1796572e4964b2f85a68bee6cf3084931f864 |
243 B | decoder.3.mlmodelc/analytics/coremldata.bin |
4832ce8a28595e3fd8f2ad4b09356a5e13d8c5039380e26c780d6e98c2123843 |
535 B | decoder.3.mlmodelc/coremldata.bin |
a72cbb218ad696ade0a810c55921b977e0c6ba4413f4e64ccfd11d38e2a063a2 |
4.3 kB | decoder.3.mlmodelc/metadata.json |
98e91a6ef82a62de16d54e2bc6892df894be523ae055b2de1ea651707c6fe8c2 |
134.0 kB | decoder.3.mlmodelc/model.mil |
8a49a8c8905661293b4caeadcc01ef020aa587e0d249114b34b14be40173a1da |
462.7 MB | decoder.3.mlmodelc/weights/weight.bin |
3cfc8912f49f6e1f6c6084454121f49ee67b4e3d37d2fd6ef0205c94baa0d0bf |
534.8 MB | embedding.safetensors |
f6cda0a14d1de23783cc69131dad6b882c2258cc8c22db840e670c04c31e1558 |
1.0 MB | rope.safetensors |
3e065a558a034185fe299917b398685c1facd0169a9eea1e629eb30c171fed81 |
9.9 MB | tokenizer.json |
Generated by tools/model_card.py from sidekick's committed data. Report a problem at https://github.com/ironblock/sidekick/issues.
Model tree for sidekickd/minicpm5-2b-coreml
Base model
openbmb/MiniCPM5-2B