Laya-MLX: A Real On-Device Alternative to Jev — Is It Really 50x Faster?

A developer who goes by mizorewww shipped something real this week: laya-mlx, an open-source Apple MLX port of Laya, a typed-decision model built by Convai Innovations. It runs entirely on-device, needs no PyTorch or cloud API, and ships with actual published benchmarks. Then, in the same breath, the announcement thread called it "50 times faster than Jev" — TypeSafe AI's own typed-decision model, which explainx.ai has covered extensively this month. One of those two claims is measured and reproducible. The other isn't in the project's own documentation at all. Here's the difference.

TL;DR

What laya-mlx actually is

Laya belongs to the same broad category explainx.ai has tracked closely through Jev: a "typed decision model" that skips free-text generation entirely and instead answers constrained questions — pick one of N choices, output a 0–1 rubric score, or return a calibrated probability — through a single encoder pass. No token-by-token decoding, no JSON parsing, no risk of the model wandering off-schema mid-generation. TypeSafe AI calls this pattern a "System One Model"; Convai Innovations built its own version independently as Laya.

laya-mlx is not that original model — it's a from-scratch runtime that ports Laya's architecture to run natively via Apple's MLX framework, avoiding PyTorch and the Hugging Face Transformers runtime entirely (tokenization still goes through HF's Rust tokenizer). That's a meaningful engineering lift on its own: rewriting a bidirectional encoder, a decision transformer, a scoring head, and an action head to run efficiently on Apple Silicon's unified memory, with no server and no network call in the loop.

The numbers that are actually published — and hold up

The repository's benchmarks, run on an M3 Max with 40 GPU cores and 128GiB of memory:

That memory figure is worth pausing on, because it's the one specific claim from the viral post that checks out cleanly: "occupies a maximum of 1G memory only on your device" lines up almost exactly with the repo's own 943.6 MiB number for the larger checkpoint. Whoever wrote that summary had actually read the benchmark output.

The claim that doesn't hold up: "50x faster than Jev"

Here's where the story gets more careful. The original announcement — a Chinese-language X post from @mizorewww, the same developer who built laya-mlx — states plainly: "Introducing a version 50 times faster than Jev, running on your device: laya-mlx!" That's a specific, falsifiable multiplier. It's also a claim that does not appear anywhere in the laya-mlx GitHub repository, its README, or its published benchmark tables. There's no comparison table against Jev, no shared task set, no note about what hardware or configuration Jev was measured on to produce the baseline the "50x" is measured against.

That distinction matters because it's the exact split explainx.ai's benchmark literacy framework exists to catch: a project can have completely real, reproducible numbers for what it actually measured (latency, memory, throughput on its own hardware) while still attaching an unsubstantiated multiplier against a competitor in marketing copy. Both things are true of laya-mlx at once. The 13.42ms figure is real. The "50x faster than Jev" figure is not something anyone outside the original post has been able to check — TypeSafe AI hasn't published a directly comparable per-decision latency number that would let a third party do that math independently.

Laya and Jev are not the same model

Worth being explicit about, since the announcement leans on the comparison: Laya and Jev are separately developed, unrelated models from different companies. Laya comes from Convai Innovations; Jev comes from TypeSafe AI, founded by Diogo Almeida, a co-inventor of the RLHF work behind ChatGPT. They land in the same conceptual category — typed decisions instead of generated text, calibrated confidence instead of raw logits — because that category is becoming a recognizable pattern in the field, not because one copied the other. A growing list of independent projects have started building on or alongside Jev specifically, and Laya sits adjacent to that broader category rather than inside it.

The solidSF quote-post: what's actually there

The @solidSF account, self-described as "foundational lab, bayesian agents, multiphysics, fully agentic hardware," quote-posted the laya-mlx announcement with a 19-second video and the caption "nobody is asking us to do these, we just love to compete." The quote-post reached 43.4K views. What the video itself demonstrates isn't something this post can verify — video content wasn't independently reviewed here, and the caption alone doesn't specify a benchmark, a product, or a result. It's included in this roundup because it's part of the public record around how the laya-mlx announcement spread, not as evidence of any additional claim.

How this fits next to Jev's actual coverage

explainx.ai has followed Jev's ecosystem closely this month: the original launch and its RLCD training method, Jev Playground and the JevBench 75.3 claim, how Jev compares to XGBoost and fine-tuned BERT classifiers, and Jev Ultrafast's 7.1-second browser agent demo. The throughline across all of that coverage has been the same one that applies here: TypeSafe AI's own disclosed numbers (20-200x faster than LLMs, 67.8% aggregate JevBench accuracy) are real, self-reported figures worth taking seriously, but every third-party multiplier claimed against Jev — including this one — needs its own independent check before it's treated as settled.

Laya-mlx earns real credit on its own terms: a working, open-source, on-device runtime with published, reproducible benchmarks and a real memory-efficiency win for anyone who wants structured decisions running locally on a Mac with no API key and no network dependency. That's a genuinely useful tool. It just isn't the same thing as a verified 50x win over a specific competitor, and treating the two as equivalent is the part of this story that doesn't survive a look at the primary source.

Honest limitations

• No independent, apples-to-apples benchmark exists comparing laya-mlx's decisions against Jev on the same task set, hardware, or accuracy metric — only laya-mlx's own solo benchmarks and TypeSafe AI's own solo benchmarks, measured separately.
• The "50x faster than Jev" figure is sourced to a single social media post by the same developer who built the project, not to the repository's documentation or benchmark suite.
• The solidSF video wasn't independently reviewed as part of this post — only its public caption and view count are cited.
• Accuracy and calibration claims for laya-mlx haven't been independently verified against a shared benchmark like JevBench; the published numbers cover latency, throughput, and memory only.

What this means for builders

If you're running structured classification or routing tasks on Apple Silicon and want to avoid a cloud API dependency, laya-mlx is worth trying on its own real merits — sub-14ms decisions and under-1GB memory are genuinely useful numbers for local, offline structured-output work. Just don't budget a migration away from Jev, or any other typed-decision model, on the strength of the "50x faster" headline alone until someone runs both on the same task set and publishes the comparison. That's a testable claim sitting right there waiting for someone to check it — explainx.ai will cover it if and when that comparison actually happens.

Related on explainx.ai

• TypeSafe AI Launches Jev: A "System One Model" That Never Hallucinates — the model laya-mlx is being compared against
• Jev vs. XGBoost and Fine-Tuned BERT: When Does the Zero-Training Pitch Win? — how Jev's own speed and accuracy claims hold up against established classifiers
• Jev Playground, JevBench 75.3: The Claims, Checked — explainx.ai's earlier fact-check of a self-reported Jev benchmark number
• Jev Ultrafast: Browser Use Puts Jev in the Browser Agent Loop — a verified, independently-timed Jev deployment for comparison
• How to Integrate Jev for Agent Routing — a practical guide to the pattern laya-mlx also implements
• How to Read AI Benchmarks Without Getting Fooled — the framework this post applies to the "50x" claim

Primary sources: mizorewww/laya-mlx on GitHub; @mizorewww on X, September 21, 2026; @solidSF quote-post on X.

This post reflects laya-mlx's published benchmarks and public announcements as of September 21, 2026. The "50x faster than Jev" comparison had not been independently verified by explainx.ai or corroborated by a published, apples-to-apples benchmark as of publication.