Modelpublic

alexm37/rune-text-v1

sourceHugging Faceapache-2.0updated 18d agoView on Hugging Face

0likes33downloads

Model Card

Rune Text v1

Rune Text v1 is the local ranking model used by Rune Keyboard Smart Typing. This artifact is the reproducible Q4KM conversion of Qwen/Qwen3-0.6B-Base revision da87bfb608c14b7cf20ba1ce41287e8de496c0cd, produced with llama.cpp revision 36b10154383b60eb15baac2c7a40d2a5f784faa7.

Artifact identity

—File: rune-text-v1-0.1.0-q4_k_m.gguf
—Bytes: 396704416
—SHA-256: 7a97111c917e19117207428971fa1c2583f2d9c2a07a6fda5b6f198b707dd9c4
—GGUF: v3, qwen3, Q4KM

Rune Keyboard verifies the exact size and SHA-256 before activation and executes inference locally in a private Android process. Typed text is not uploaded by the keyboard.

Smart Typing evidence

The frozen production-candidate holdout at the user-selected 95% point-precision profile produced:

Language	Automatic changes	Correct	Precision	False changes / 1000
English	304	293	96.38%	0
Russian	359	342	95.26%	1
Spanish	408	396	97.06%	0

Original remained available in every candidate set. These measurements use the bounded Rune scorer and frozen product policy. The complete physical Fold latency, thermal, cancellation and process-death qualification is still pending, so this repository is marked experimental rather than a final Rune Keyboard release.

Intended use

The model scores a bounded list of spelling or punctuation continuations. It is not exposed as an unrestricted text generator. Rune Keyboard retains protected-token checks, bounded candidate search, confidence thresholds, immediate Undo, and local deterministic capitalization for supported proper names.

License

Apache License 2.0. See LICENSE.