laya-mlx

Laya on Apple MLX. Under 1GB RAM, 7 to 14ms on M3 Max. Snake at 60 decisions a second, no PyTorch, no cloud.

by @mizorewwwAdded Sep 22, 2026

介绍比Jev快50倍，在你设备上跑的laya-mlx! 只在你的设备上占用最高1G内存 Laya是一个开源的类似于Jev的,基于文本输出概率的分类系统我将其移植到MLX,并且做了一些性能优化! 视频中就是这个模型在我的本地M3Max上玩贪吃蛇这个模型能够以每秒决策60次的速度玩贪吃蛇！ github.com/mizorewww/laya…

Watch on X

12:48 AM · Sep 20, 2026

14.0K

Read 295 replies

Visit project View post on X

Related builds

Oliver Yu

@oliver_yu9

·Follow

Everyone’s making Laya vs Jev game demos, so I tried a different experiment: Can the same Laya model run on both the MLX GPU and Apple Neural Engine? Then I used both at once for serving. M4 Max, bursty load: P99 1592.9 ms → 79.5 ms github.com/tc3oliver/laya…

Watch on X

3:46 AM · Sep 24, 2026

laya-apple

Same Laya weights on MLX GPU and Apple Neural Engine at once. P99 1593ms to 80ms on M4 Max.

Visit project

Hemant

@heman10x

·Follow

Did my openJev Verdict(Oss jev made on just 151M param base model) hit #2(possibly #1 even?) on JevBenchmarks> When JevBench v1.2 evaluated openJev Verdict earlier this week, we were sitting at #11 with a score of 66.2. I audited the benchmark run this morning and realized the Show more

Hemant

@heman10x

Did my OSS-Jev called Verdict, beat Official Typesafeai 's Jev ?, and Laya (the best performing OSS Jev) on their own benchmark? When I first tested Verdict 1.0(my earlier attempt of oss-jev) earlier this week, it scored 26.10% accuracy. Uniform random guessing on that split is

3:25 PM · Sep 20, 2026

Read 1 reply

Xjevopen source

Sep 22, 2026

openJev Verdict

Open 151M decision model on ModernBERT. JevBench #2 on public tasks after the inference-engine fix. Weights and a WebGPU playground.

Visit project

Nandakishor m

@Nandakishorm1

·Follow

Asking everyone to fine tune github.com/NandhaKishorM/… Just use any kinda agentic harness like claude code, codex etc. Change the code and use free kaggle GPUs to train your own use cases. Base model has limitations

10:21 AM · Sep 21, 2026

727

Read 37 replies

Xjevopen source

Sep 22, 2026

Laya

Open Apache 2.0 System One model. ModernBERT, typed choice/score/noul in one pass. Fine-tune notebook on Kaggle T4s. Not Jev.

Visit project

Jared Palmer

@jaredpalmer

·Follow

Kev has now been refactored on top of Qwen3.5. New checkpoints are now available at 0.8B, 4B, and 9B along with a fine-tuning script you can use with @modal. github.com/jaredpalmer/kev

Jared Palmer

@jaredpalmer

UPDATE: Kev-0.6B, 4B, and 8B are now available. Kev is a family of small open source Jev-like decision models you can train and run yourself. This new family is based on Qwen3 using the same LoRA + small pointer head technique as before, but scaled up. Out of domain, on data

2:52 PM · Sep 21, 2026

752

Read 34 replies

Xjevopen source

Sep 22, 2026

Kev

Jared Palmer's open Jev-like family on Qwen3.5. 0.8B, 4B, 9B. TypeSafe SDK with a base_url change. Train on one H100.

Visit project