Downloads · 30 days
0
MaGuN4U/Kanana-Longing
Kanana-Longing is a machine learning model from MaGuN4U. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as other.
Powered by Kanana · Linux x8664 · binary prototype · black-box evaluation encouraged
Downloads · 30 days
0
Access
Public
Updated Sep 17, 2026
Repo size
11.4 MB
Likes
0
Public
Click a slice to open those files.
.zip11.4 MB · 100%
From the Hugging Face model README
Powered by Kanana · Linux x86_64 · binary prototype · black-box evaluation encouraged
A 9.53 MiB Korean language cell is not supposed to look competitive with a 1.3B reference model. By itself, it does not. The interesting part is what happens when it is embedded in the complete local system.
Kanana-Longing is an independent experimental prototype built around a Kanana-derived language cell and a proprietary local runtime. This repository distributes the qualified Linux x86_64 binary release; it is not a Transformers checkpoint and the proprietary runtime implementation is intentionally not published.
The frozen language-cell artifact inside this release is 9,993,551 bytes (9.53 MiB). The current qualified binary ZIP is 11,361,360 bytes (10.84 MiB). Host runtime dependencies are not bundled in those figures.
Published black-box evidence shows a deliberately mixed result: the small language cell is materially weaker than the Kanana2 1.3B reference on intrinsic/factual tests, while the complete local system performs much better on one frozen structured/semantic system suite.
| Published evaluation | Kanana-Longing | Kanana2 1.3B reference |
|---|---|---|
| Korean semantic pair-ranking 64 — language cell only | 25/64 (39.06%) | 38/64 (59.38%) |
| Closed-book external 17 — language cell only | 5/17 (29.41%) | 11/17 (64.71%) |
| Frozen structured/semantic system suite 70 — end-to-end | 60/70 (85.71%) | 41/70 (58.57%) |
| Fresh symmetric marker suite 9 — end-to-end | 3/9 (33.33%) | 6/9 (66.67%) |
| Integrated factual suite 100 — end-to-end | 31/100 (31%) | 57/100 (57%) |
Interpretation matters: this is not evidence that a 9.53 MiB standalone model generally beats Kanana2 1.3B. It does not. It is evidence that the complete prototype has a substantially different capability profile from the language cell alone on the published test suites.
hf download MaGuN4U/Kanana-Longing \
Kanana-Longing-Prototype-v0.2.3-linux-x86_64.zip \
Kanana-Longing-Prototype-v0.2.3-linux-x86_64.zip.sha256 \
--local-dir .
sha256sum -c Kanana-Longing-Prototype-v0.2.3-linux-x86_64.zip.sha256
unzip Kanana-Longing-Prototype-v0.2.3-linux-x86_64.zip
./coreana
SHA-256
bffad17ea9502c21438334556c02a3102583827a6bf3a0a73b969d99e4d3cbbb
The release expects a compatible Linux x86_64 host and uses host-provided runtime libraries. The packaged README lists glibc, libstdc++, libgcc_s, SQLite3, OpenSSL libcrypto, ICU, zlib, zstd and Python 3.13 among the runtime dependencies.
The public release manifest records the following qualification results:
These checks qualify the distributed artifact; they are not claims of general model safety, security, or benchmark superiority.
This prototype is intentionally suitable for black-box testing. Useful independent reports include:
If you test it, negative results are as useful as positive ones.
Kanana-Longing contains a Kanana-derived language cell. Use and redistribution of the Kanana-derived portions are subject to the Kanana Open License Agreement. The release ZIP contains the applicable Kanana license and third-party notices, including SentencePiece/Apache-2.0 and bundled native dependency notices.
Kanana-Longing is an independent prototype and is not presented as an official Kakao product or as endorsed by Kakao.