Maintainer of this skill?
Automated security scan — not a certification.
by youssofal
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.
# Add to your Claude Code skills
git clone https://github.com/youssofal/MTPLXLast scanned: 9/17/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-09-17T09:13:42.636Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}See how MTPLX compares with popular alternatives.
MTPLX is an open-source testing skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by youssofal. 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. It has 2,657 GitHub stars.
Yes. MTPLX passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/youssofal/MTPLX" and add it to your Claude Code skills directory (see the Installation section above).
MTPLX is primarily written in Python. It is open-source under youssofal on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other Testing skills you can browse and compare side by side. Open the Testing category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh MTPLX against similar tools.
No comments yet. Be the first to share your thoughts!
Top skills in this category by stars
From $49 · 30 days · terms & refunds
⚠️ Third-Party Software Notice
This skill is third-party open-source software developed and hosted independently on GitHub. SkillsLLM is an informational directory and does not control or maintain the underlying repository.
Any security checks, ratings, or warnings displayed by SkillsLLM are automated and limited in scope. They do not constitute a security certification or guarantee that the software is safe, error-free, or free from malicious code, vulnerabilities, compromised dependencies, or prompt-injection risks.
Review the source code, permissions, dependencies, and configuration before installing or running any third-party skill. Use is at your own risk. To the maximum extent permitted by applicable law, SkillsLLM is not liable for losses arising from third-party software.
Automated security scan — not a certification.
MTPLX is a native Mac app and a command line that runs local language models on Apple Silicon with the model's own multi-token prediction (MTP) heads. It runs Qwen 3.8 Flash Next, the 125B mixture of experts, Qwen 3.8 27B and Prism ML's Ternary Bonsai 2 27B, plus Qwen 3.6, Qwen 3.5 and Gemma 4. The model drafts several tokens ahead of itself, one batched forward pass verifies the draft, and tokens are committed through exact rejection sampling with residual correction. The sampler works at the model's native temperature, and decode runs at around twice the speed of plain decoding: measured 1.6x on a 16 GB M4 Mac mini and 2.24x on an M5 Max.
Every number below was measured on a MacBook Pro M5 Max with 128 GB, fans verified at maximum, sampled at the model's own settings. Conditions and sources for every row, with the raw logs where they exist, are at mtplx.com/benchmarks.
| Model | Speed | Run |
|---|---|---|
| Qwen 3.8 Flash Next, Optimized Speed | 63.5 tok/s | 65,502-token prompt read at 1,094 tok/s (768 on 2.11.3), first token after 60.1 s instead of 85.5 s, 512 tokens generated, thinking off, MTPLX 2.12.0 (56.1 tok/s on 2.11.3) |
| Ternary Bonsai 2 27B, Optimized Speed | 64.4 tok/s | 4,061-token prompt, 512 tokens generated, MTP depth 1, thinking off, MTPLX 2.12.0 (52.6 for the 4-bit Qwen 3.8 27B in the same session), 11.4 GB peak |
| Qwen 3.8 Flash Next, Optimized Speed | 125.8 tok/s | one OpenCode request: 1,301 tokens generated, 18,539-token prompt with 18,364 tokens served from cache, MTP depth 3, MTPLX 2.11.3, 16 September 2026 |
| Qwen 3.8 Flash Next | 79.3 tok/s | 9k-token code prompt, 1,500 tokens generated, thinking off, two alternating boots each, MTPLX 2.11.3 (62.5 on 2.11.2) |
| Qwen 3.8 Flash Next | 61.8 tok/s | 109k-token OpenCode turn, mean of two runs, MTPLX 2.11.3 |
| Qwen 3.8 Flash Next | 50.3 tok/s | 200k-token OpenCode turn, warm, mean of two runs, MTPLX 2.11.3 |
| Qwen 3.8 27B, Optimized Speed | 87.6 tok/s | rewriting a file it just wrote, stock settings, MTPLX 2.10.0 |
| Qwen 3.8 27B, Bare Speed | 65.2 tok/s | fresh coding task at official Qwen 3.8 sampling, generation to the model's own stop, MTPLX 2.7.0 |
| Qwen 3.6 27B, Optimized Speed | 81.74 tok/s | the 27B record on a fresh generation: 192-token bench, thinking off, temperature 0.6, twin runs, 2.69x over 30.37 plain decode, 2 July 2026, raw logs |
| Qwen 3.5 4B, Optimized Speed | 227.8 tok/s | depth 3, 1.71x over 133.6 plain decode, MTPLX 2.2.0 |
MTPLX uses rejection sampling with residual correction at the model's native temperature. For 2.11.3, MTPLX compared a thousand four-token samples from the fast path with a thousand from the plain path at temperature 1, top-p 0.95 and top-k 20, by token id. Those measurements matched within the plain path's own noise on both Flash Next and the 27B Quality pack. From 2.12.0, image requests also use the compiled verifier, and that sampled comparison has not yet been repeated for images. The acceptance rule follows the Leviathan and Chen rejection sampling theorem; numerical equivalence of the model's execution paths is checked separately. There is no second draft model eating your RAM, and no greedy shortcut that quietly changes what the model would have said.
How MTPLX compares with mlx-serve, oMLX, LM Studio, Ollama, llama.cpp and mlx-lm, with a version, a machine and a date on every number: mtplx.com/compare.
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 mtplx on your PATH, and then measures your machine to pick the fastest decoding depth.
Recommended for coding: Qwen 3.8 27B Optimized Speed is a 4-bit dynamic quant with great coding speeds and good quality. Its two siblings sit right under it in the app and CLI: Bare Speed (quickest burst chat speeds, lower quality and slower on long coding tasks) and Optimized Quality (8-bit dynamic quant, the highest fidelity of the three). Qwen 3.8 Flash Next is offered as Bare Speed from 96 GB and as Optimized Speed from 128 GB, and on a Mac with 256 GB or more Flash Next Optimized Speed is the first recommendation. Qwen 3.6 Optimized Speed V2 remains available below them.
The CLI on its own:
brew install youssofal/mtplx/mtplx
mtplx start
or python3 -m pip install mtplx if you prefer pip. All releases are listed at mtplx.com/releases.
Requirements: Apple Silicon (M1 or newer), macOS 14+. The app and CLI use one recommendation policy on M3, M4 and M5: Qwen 3.5 4B Optimized Speed below 16 GB, Ternary Bonsai 2 27B from 16 GB, Qwen 3.8 27B Optimized Speed from 32 GB, and Qwen 3.8 Flash Next Optimized Speed from 256 GB, with Flash Next Optimized Quality second. M1 and M2 keep their FP16 recommendations. Installed models and explicit selections remain available.
Ternary Bonsai 2 27B and Flash Next Optimized Quality require MTPLX 2.12.0 or later.
Qwen 3.8 Flash Next is Qwen's 125B-A6B preview of the Qwen4 architecture: a hybrid GatedDeltaNet mixture of experts with Qwen Sparse Attention and a 51B-parameter n-gram table. MTPLX 2.10.0 was the first Apple Silicon backend for the family, and it runs the model's own MTP head as an exact speculative decoder. Bare Speed is offered from 96 GB of unified memory and Optimized Speed from 128 GB. MTPLX 2.12.0 adds Optimized Quality for Macs with 256 GB or more:
Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed: dynamic 4-bit with the sparse-attention projections at 8-bit. 115.1 GB download including the 32 GB n-gram table, about 83 GB resident.Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed: flat 4-bit, the quickest build. 106.3 GB download, about 74 GB resident.Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Quality: 8-bit group-64 body and MTP head, BF16 structural tensors, and a 4-bit group-32 n-gram table. 169.96 GB download (169,958,537,520 bytes); the weights need about 128.5 GiB. It has not yet been run on a 256 GB Mac, and its speed has not been measured.The n-gram table streams from SSD on every Mac, so only the weights stay in memory. Context window 262,144 tokens; 261,120-token prompts decode. Image input works. In the app, pick "Qwen 3.8 Flash-Next Optimized Speed"; from the terminal:
mtplx serve --model Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed
Then point OpenCode, Pi, Hermes, Claude Code, Cline, Cursor or anything that speaks the OpenAI or Anthropic API at http://127.0.0.1:8000. The guide with every measured number and its conditions: mtplx.com/models/qwen3.8-flash-next.
Youssofal/Ternary-Bonsai-2-27B-MTPLX-Optimized-Speed is a compact ternary 27B
model from Prism ML with vision and MTP. It is the first recommendation on
M3, M4 and M5 Macs with 16 to 31 GB. The old Bonsai-3.8-27B-MTPLX-Optimized-Speed folder and
mtplx-bonsai-38-27b-optimized-speed served ID resolve to the same entry;
clients use mtplx-bonsai-2-27b-optimized-speed.
The pack is 8.85 GB with its vision tower and the Qwen3.8-27B MTP head, which
defaults to depth 1. Reasoning effort is medium by default; xhigh is also
available. On an M5 Max it decodes 64.4 tok/s at 4K and 57.1 at 16K, against
52.6 and 51.0 for the 4-bit 27B, in about half the memory.
On a 16 GB Mac it gets an 8,192-token window, with a measured peak of
11.80 GiB under a 16 GB memory budget; agent clients such as OpenCode and
Hermes need 18 GB or more.
Qwen 3.8 27B is the coding flagship for Macs with 32 GB or more. MTPLX shipped it on 15 August 2026, the day after Qwen released it. Three packs, each with an FP16 sibling for M1 and M2 that the app and CLI pick automatically:
Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed: 4-bit dynamic, 20.4 GB, 23.6 GB peak. The default for coding. Against the bf16 model on a mixed corpus of code, prose and JSON it agrees on 96.0 percent of top-1 tokens with a KL divergence of 0.012.Youssofal/Qwen3.8-27B-MTPLX-Bare-Speed: flat 4-bit, 16.0 GB, 17.0 GB peak. Quickest chat speeds.Youssofal/Qwen3.8-27B-MTPLX-Optimized-Quality: 8-bit dynamic, 29.4 GB, 32.7 GB peak, for 36 GB or more. 99.3 percent top-1 agreement with bf16, KL 0.0005.The guide: mtplx.com/models/qwen3.8-27b.
The catalog lists MTPLX packs under Youssofal. For measured packs, "Fits" is the catalog peak while serving, next to the smallest Mac the app and CLI will offer it on. The preset column is what MTPLX resolves by itself, so this table is what you get by doing nothing.
Model (Youssofal/...) |
Fits | What it is for | Preset |
|---|---|---|---|
Qwen3.5-4B-MTPLX-Optimized-Speed |
8 GB and up, peaks at 2.9 GiB | 4-bit. The fastest fit for smaller Macs. | Sustained, depth 3 |
| `Qwen3.5-4B-MTPLX-Optimized-Quali |