Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf praxis-app/praxis-models:Q4_K_M# Run inference directly in the terminal:
llama cli -hf praxis-app/praxis-models:Q4_K_MBuild from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf praxis-app/praxis-models:Q4_K_M# Run inference directly in the terminal:
./build/bin/llama cli -hf praxis-app/praxis-models:Q4_K_MUse Docker
docker model run hf.co/praxis-app/praxis-models:Q4_K_Mpraxis-models
Model assets redistributed for the Praxis language-learning app.
This repository contains no original work. Every file here is a byte-for-byte copy of an upstream release, re-hosted only so the app has a stable, version-pinned download source. Nothing has been re-trained, re-quantized, or otherwise modified.
Why re-host at all?
The app downloads models on demand and verifies them with a SHA-256 that is compiled into the client. Pointing directly at an upstream repository would mean that any upstream change — a re-export, a file rename, a repository removal — silently breaks downloads for every already-shipped client. Re-hosting gives us a path we control; upstream is kept as a fallback mirror.
Contents
sense-voice/ — SenseVoice ASR (int8)
| File | Size (bytes) | SHA-256 |
|---|---|---|
model.int8.onnx |
239,233,841 | c71f0ce00bec95b07744e116345e33d8cbbe08cef896382cf907bf4b51a2cd51 |
tokens.txt |
315,894 | f449eb28dc567533d7fa59be34e2abca8784f771850c78a47fb731a31429a1dc |
Source: csukuangfj/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17
— ONNX export of SenseVoice by FunAudioLLM,
exported for sherpa-onnx by Fangjun Kuang.
The SHA-256 of model.int8.onnx matches the upstream LFS pointer exactly, which is
the check we use to confirm these are unmodified copies.
Languages: Chinese, English, Japanese, Korean, Cantonese.
mt-qwen3-1.7b/ — Qwen3-1.7B machine translation (GGUF, Q4_K_M)
| File | Size (bytes) | SHA-256 |
|---|---|---|
Qwen3-1.7B-Q4_K_M.gguf |
1,107,409,472 | b139949c5bd74937ad8ed8c8cf3d9ffb1e99c866c823204dc42c0d91fa181897 |
Source: unsloth/Qwen3-1.7B-GGUF
(revision d7f544eead698dbd1f15126ef60b45a1e1933222, file Qwen3-1.7B-Q4_K_M.gguf) —
GGUF quantization of Qwen3-1.7B by the Qwen team,
quantized by Unsloth. Runs via llama.cpp.
The SHA-256 matches the upstream LFS pointer exactly.
Used for on-device translation so the app works offline and without per-request API cost.
Verifying
sha256sum model.int8.onnx tokens.txt
Any mismatch means the file was corrupted in transit or tampered with — do not use it.
License
Apache-2.0, inherited from the upstream releases. All credit for these models belongs to their original authors, cited above.
- Downloads last month
- 26
4-bit
Install (macOS, Linux)
# Start a local OpenAI-compatible server with a web UI: llama serve -hf praxis-app/praxis-models:Q4_K_M# Run inference directly in the terminal: llama cli -hf praxis-app/praxis-models:Q4_K_M