Instructions to use ryoshu/qwen-mtp-chat-runtime with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use ryoshu/qwen-mtp-chat-runtime with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir qwen-mtp-chat-runtime ryoshu/qwen-mtp-chat-runtime
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
qwen-mtp-chat-runtime
Compiled Swift/Metal runtime binaries for
ryoshu/qwen-mtp-chat โ a local
chat CLI for Qwen 3.8 27B with native multi-token-prediction (MTP)
speculative decoding on Apple Silicon.
This is not a model repository. It contains no model weights โ only the compiled runtime that loads and drives them. Model weights are hosted separately by their own publisher; see Provenance below.
Contents
| File | What it is |
|---|---|
mlxfast-swift |
trusted CLI: protocol/timing/gate driver, links no MLX |
mlxfast-runtime-worker |
sandboxed worker that loads the model and runs the forward pass |
mlx.metallib |
compiled MLX Metal kernels the worker dispatches |
Requirements: Apple Silicon Mac, macOS 14+. No Xcode/Swift toolchain needed to run these โ they only link system frameworks and the OS Swift runtime.
Provenance
Compiled from Layr-Labs/qwen-3.8-mtp-challenge's
MIT-licensed trusted harness (Sources/MLXFastCLI,
Sources/MLXFastTrustedHarness, Sources/MLXFastCore,
Sources/MLXFastModel, Sources/MLXFastTransform) and its vendored MLX
Swift stack (mlx-swift, mlx-swift-lm, swift-transformers; MIT/
Apache-2.0), plus one addition: a chat subcommand for real, streaming,
arbitrary-prompt generation over the native-MTP block-decode session,
written for ryoshu/qwen-mtp-chat
and not part of the upstream challenge's benchmark/scoring surface.
Not an official artifact of, or endorsed by, Layr Labs, Inc., EigenLabs, or
the mlxfast challenge organizers. Full attribution:
NOTICE.md
in the companion GitHub repo.
Model weights (downloaded separately, not hosted here)
install.sh in the companion repo fetches these directly from their
publisher:
- Backbone:
EigenLabs/Qwen3.8-27B-4bit@eda45ab47f465d08d6558f0353a2346e2eb9d5b3 - MTP head:
EigenLabs/Qwen3.8-27B-MTP-bf16@26a328e070875b0314d652a039b6b59902690f03
Usage
Don't download these files directly โ use the installer, which fetches and sha256-verifies everything:
git clone https://github.com/ryoshu/qwen-mtp-chat.git
cd qwen-mtp-chat
./install.sh
./chat.sh --prompt "your question"
License
MIT, inherited from the upstream trusted harness โ see the companion repo's
LICENSE and
NOTICE.md.