mtplx
UNCLAIMEDINFERENCE ENGINESEDGE AND ON-DEVICE
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. Listed under Inference engines in the Inference / Serving layer. 2 of 11 comparison attributes are recorded, each with the page it was read from.
A native Mac app for local LLMs. Qwen 3.8 Flash Next at 125 tok/s on an OpenCode request, Qwen 3.8 27B up to 87.6 tok/s, the 81.74 tok/s record on a 27B, all on an M5 Max. Twice the speed of plain decoding, output exact at any temperature.
4 OF 4 SIGNALS HELD
What it runs on, and what runs on it
0 cited edgesComputed from the stack graph, both directions. Every cited edge links the page it was read from; a vendor-added one links the vendor’s own source — this is the reason to be here rather than on the vendor’s own site.
No edges on record for mtplx.
Nothing in the directory documents depending on it, and its own docs do not name what it runs on. An edge is only stored when a source states it.
Submit a documented integration →From mtplx? No claim is needed for this: a link to your docs that names what mtplx runs on is enough.
Open-source signals
A HOLLOW RING IS UNMEASURED, NOT ZERO · WE DO NOT SUBSTITUTE A PARENT COMPANY’S REPOSITORY FOR A PRODUCT THAT HAS NONE
Price
We hold no price for this product.
This is open-source software: you can run it yourself and the licence costs nothing. If the project also sells a hosted service, we do not yet hold the page it prices it on. If that changes — or if you publish /.well-known/pricing.json — it appears here with a capture date.
How we read prices →What its categories ask
2 answered · 9 unanswered11 questions declared by Inference engines and Edge and on-device. Most rows across the directory read like these.
“Forge converts it to MLX and measures the speedup on your Mac.”2026-10-20
A ROW WITH NO ANSWER IS NOT A ROW WITH A “NO” · A VENDOR MAY ANSWER A BLANK ONE, MARKED VENDOR CLAIMED WITH ITS SOURCE · A DOCUMENTED VALUE IS DISPUTED, NOT OVERWRITTEN, AND STAYS VISIBLE WHILE WE RE-READ THE SOURCE
Practitioner reviews
Nobody has written one yet.
A review here is structured, not a star: hardware, scale band, what broke in production, and whether you would choose it again.
NO PERCENTAGE IS SHOWN UNDER FIVE REVIEWS · NO STAR RATINGS, EVER
Also in this category
67 products in Inference engines. Comparison defaults to within the category — pairing this with a GPU cloud would produce a page that means nothing.
Frequently asked
Is mtplx open source?
What does a blank mean on the mtplx page?
Where do the mtplx figures come from?
Can mtplx change what this page says?
Something here is wrong — how do I fix it?
Is this listing run by mtplx?
Cite this page
A model that quotes us should quote us correctly. The reading and its date sit inside the citation rather than beside it, because a quotation carrying only a URL goes wrong the week the URL changes.
infratoapp, “mtplx — Inference engines”, read 20 October 2026. Hardware: Apple Silicon as published at mtplx.com. https://www.infratoapp.com/products/mtplx
What mtplx can and cannot change here
Claiming the organisation lets its admins and editors change the tagline, description, logo, docs link and deployment model of every product that org owns — live, labelled as theirs, and logged. Answers to blank questions and dependencies go live too, marked as the owner’s and sourced; a documented answer cannot be overwritten. Everything else on this page is a stated fact with a dispute route: price, licence and OSS figures. Those columns are ungranted at the database level, so they are not a feature we withheld — they are a write that cannot happen. Name, category, website, pricing page and package names can be changed by the owner only through our review. A dispute never deletes a value; it marks it and publishes the correction.