youssofal/MTPLX
The fastest way to run Qwen 3.8 Flash Next, Qwen 3.8 27B and Ternary Bonsai 2 27B on a Mac: 125 tok/s in OpenCode on an M5 Max, and a 27B model on 16 GB Macs. Native MTP speculative decoding on Apple Silicon, exact at any temperature. OpenAI and Anthropic compatible local server.
Build with Backblaze B2
SDKs, agent skills, IDE extensions, and reference pipelines from Backblaze Labs. All open source.
Loading star history...
Where it stands
Rank, growth pace and release cadence relative to everything GStars tracks.
5 releases in the last 90 days, the latest 34 days ago, with 44 contributors on GitHub.
Peaked at #33 in the weekly ranking the week of Sep 6, its best in the past year.
Use Cases & Benefits
- Runs local large language models on Apple Silicon Macs using native multi-token prediction for faster exact decoding.
- Delivers up to twice the speed of traditional decoding by leveraging native MTP speculative decoding and exact rejection sampling without quality loss.
- Use for running Qwen 3.8 and similar models locally on macOS with optimized speed and memory usage for coding tasks.
- Use for serving OpenAI- and Anthropic-compatible APIs locally to integrate with chatbots, agents, or custom AI applications.
- Use for building and verifying custom MTP models from Hugging Face checkpoints with the Forge tool to optimize inference speed.
Should you depend on it?
License terms, project hygiene and risks, from GitHub metadata and GStars' snapshots.
License
Apache 2.0
Permissive: commercial use, modification and redistribution are allowed with attribution; includes an explicit patent grant from contributors.
Community files
- README: present
- Contributing guide: present
- Code of conduct: present
Watch-outs
No watch-outs from GStars' checks.
General guidance from repository metadata, not legal advice. How we score
From the README
MTPLX is a native Mac app and a command line for running local language models with multi-token prediction. Modern models like Qwen 3.5/3.6/3.8 ship with built-in MTP heads. Almost no runtime uses them. MTPLX does: the model drafts several tokens ahead of itself, verifies each drafted block in a single batched forward pass, and commits tokens through exact rejection sampling with residual correction. Same model, same output distribution, measured 1.6x faster on a 16 GB M4 Mac mini and 2.24x on an M5 Max.
There is no second draft model eating your RAM, and no greedy shortcut that quietly changes what the model would have said at real sampling settings. The acceptance math is the Leviathan and Chen rejection sampling theorem with residual correction, so
temperature=0.6, top_p=0.95behaves exactly like normal decoding, just faster.The Mac app is the easiest way in. Download the DMG at mtplx.com, drag it to Applications, and the app takes care of everything else: it checks your hardware, recommends a model that actually fits your memory, downloads it, sets up its own Python engine (no Homebrew needed), installs fan control, puts
mtplxon your PATH, and then measures your machine to pick the fastest decoding depth.
Daily snapshots
Stars gained per day, and how forks and open issues moved, from GStars' daily snapshots
+1,250 stars over 79 days across 31 snapshot days; best day +67 on Sep 2.
Forks
210+109
Open issues
116+62
Compared with peers in Backend
Fastest-growing repositories GStars tracks in this cohort for the week of Nov 1, by 7-day star growth
| Repository | Stars | +7d | Health | Last release |
|---|---|---|---|---|
| anthropics/skills | 187.8K | +2.9K | 15 | — |
| DuarteSantos8/openGym | 19.7K | +2.4K | 60 | 2 days ago |
| shiyu-coder/Kronos | 47.0K | +2.3K | 16 | — |
| NationalSecurityAgency/ghidra | 98.2K | +2.3K | 64 | 14 days ago |
| microsoft/VibeVoice | 56.0K | +1.1K | 29 | — |
| youssofal/MTPLX | 2.7K | — | 60 | 34 days ago |
| Category median | 56.0K | +2.3K | 29 | — |
Discover Repositories
Search across tracked repositories by name or description