Seen on X
Demos, games, ports and benchmarks posted on X, picked by hand. Each one links back to the original post and, where there is one, to the project in the directory.
Just implemented Laya-MLX into one of my projects. Previously, I relied on the Sightengine API for reasoning and classification, which introduced external dependencies, latency, and data leaving the local environment. I’ve now moved the decision-making layer to Laya-MLX, making classification faster, more efficient, and private by running it locally. Special thanks to @Nandakishorm1. repo :-
Project: SafeSphere.ai→The Laya decision model, now in the AI SDK. Same API as jev. Runs locally on MLX. ~16ms per decision ~2GB memory No more Python glue. uv tool install it, call it from TypeScript.
Project: laya-mlx-http→5/5 It's a living benchmark: I'll keep adding models as they ship. Next up: Fastino GLiDE, Laya Multilingual, and plain chat-model baselines. Results, charts and method: Code and raw data:
Project: system-one-bench→A local reasoning model on an Apple Silicon, with no server Laya on Mac M4 CoreML: - Runs fully offline through the CoreML neural engine, weights on Hugging Face - Four commands to install: uv init, add laya-coreml, download the ANE model, run - Snake demo included so you can watch it decide on every frame The Mac is the inference box.
UPI complaint triage using Laya mlx Here’s what it does : routes and analyses queries on your UPI app based on input text. Laya triage - route clean cases. - contradictions are sent to a reviewer Would love to know your thoughts! #layamlx #ai
🚀 Excited to share the first decision models for Lithuanian medical texts! By our @AIML_VU we’ve released three open models for calibrated typed decisions on Lithuanian MEDLINE abstracts: • Qwen3.5-4B-LT-med • medgemma-1.5-4b-it-decitions-lt-med • laya-multilingual-lt-med Trained/adapted on MEDLINE-S1-LT for study design, disease area, humans/animals, adverse effects, drug therapy, age groups etc. Strong accuracy + calibration on the held-out Lithuanian test set. Research models - not medical devices or clinical advice. Links:
Same three primitives: choice, score, yes/no. Up to 64 questions per request, one forward pass, a probability distribution and a confidence on every answer. A four-question ticket triage is 345 tokens, or $0.000007.
Laya makes decisions instead of writing text. Use a few hundred labeled examples to set thresholds before relying on it. The playground has four presets. Docs: Model by Convai Innovations:
Project: laya→
testing Laya "ms_per_question":"31.0ms" nice!
Project: laya→Spent a morning testing a small open-weights local model (Laya-MLX, fully offline on a MacBook Air, $0 per call) on real tasks: inbox triage, tagging 100 SAP decks, categorizing blog posts, and a prefilter pattern. Honest numbers in this thread.
All measured on an M4 Air, 16GB. Local model: 656MB RAM, tens of ms per decision, $0. Same rules as always: measure on your own data, engineer the input before blaming the model, never trust a confidence gate alone.
关于执行速度：我用 1000 条数据按照 40 个标签分类。 Laya coreml 平均耗时 70.7ms 总耗时 66.4s Laya mlx 平均耗时 221.1ms 总耗时 207.6s AgentJev0.6b 平均耗时 4s 总耗时 21m34s 并且执行失败
Do check out this public repo ( by @m37labs for those interested in understanding Laya-MLX, a new kind of encoder-only architecture and its application. Emerging systems like Jev, Laya can serve as a `System 1` brain in different implementations.
Project: laya-mlx-per-turn-classifier→Laya: Apache 2.0 open weights (ModernBERT / mmBERT family). Run it yourself. Authors report ~33 ms for one question on a T4. Multilingual checkpoint + a Router that picks the checkpoint by language.
Project: laya→
I used laya-mlx to play Tetris. The local model is incredibly fast — with GPU acceleration, it returns results in around 20ms.
Project: laya-mlx→
💻 mizorewww/laya-mlx by @mizorewww Runs Laya typed decisions locally in MLX for choice, score, and boolean outputs. Keeps upstream weights, formatting, calibration, and schema. Key points: • 7.39 ms P50, 395.0 q/s for multilingual 322M on M3 Max • 13.42 ms P50, 146.8 q/s for English 421M on M3 Max • Matched upstream selected answers on 63/63 validation questions in FP16/FP32 Shows that bidirectional decision models can replace token generation for low-latency structured local inference on Apple Silicon. ⭐ 4.2k stars
Jev's open-source alternative Laya-MLX runs on Apple Silicon with 1GB RAM and 7ms decisions. Could local decision models replace cloud LLMs for structured tasks? #Jev #Laya #LayaMLX #AppleSilicon #LLMs
Laya is a 421M-parameter encoder that answers typed questions about a text or JSON state in one ~33 ms forward pass on a T4. No generation, so nothing to parse. Runs locally. Model: Our read + 2-min lesson:
Project: laya→laya-coreml runs categorical decisions on Apple Neural Engine in 5ms P95. M3 Max sustained 49 decisions/sec in a Snake game demo. 2.78× better energy than MLX. No PyTorch, no token generation—just deterministic multiple-choice answers for game AI and real-time policy loops. #coreml #applesilicon #inference
Merged into Laya (20k+ stars). 3 days ago: laya-multilingual never picks the first score option (0/300 ja, 0/290 en). Today: 0.3.7 ships a Limits note, the maintainer named me the retrain's before/after tester, and merged my label-free regression check.
Project: laya→@ROBO_F8 **Short explainer:** Python coding-agent harness that places a fast local System-1 decision model ([Laya]( in front of any worker model (local or paid). Laya scores, routes, filters and curates; the worker only generates.
Project: laya→
四个名字我第一次看成四款聊天模型 Jev 是 TypeSafe 云上的决策 分类、路由、打分、是或否 走 API Laya 把权重放出来 能下到本地，也能按自己的任务微调 Kev 用 Qwen3.5 做成相近的开源款 要研究和自己训，看它 Laya-MLX 是 Laya 在 Mac 上的跑法 模型还是那份，推理换一条路 我选的时候只问 数据能不能出门 能出门用 Jev 留在家里用 Laya 要改权重再看 Kev
Laya runs typed decision models from Node.js and TypeScript through ONNX Runtime, without Python or PyTorch. It returns probabilities for fixed choices, score distributions, and yes/no results in the Python reference format, batches questions into one model run, and requires Node.js 20+.
Project: laya→Laya-CoreML runs Laya-style typed decisions on Apple Silicon through Core ML and the Neural Engine. It returns probabilities and scores instead of generated text, includes offline Python and command-line tools, and supports macOS 15+ without PyTorch, Transformers, or MLX at inference time.
Project: laya-coreml→
TypeSafe's Jev blew up this week: a model that skips text generation entirely, returns typed decisions with probabilities instead. Days later, someone open sourced Laya, ported to run natively on Apple Silicon. I ran it myself.
Jalankan model keputusan Laya langsung di Apple Silicon via MLX — keputusan 7–14 ms per request, tanpa PyTorch atau API cloud.
Project: laya-mlx→
This is the one from HF. Laya is bottom left, but Decider 2B is middle of the pack, and they actually both consume about 2GB of VRAM - Decider 2B is actually smaller in VRAM (1.4GB vs 2.3 GB) runtime - at least how I'm serving it. Decider one-shotted same or higher compared to my post-tuned (5 rounds of tuning). Results: Battery head-to-head (same cases, same weights/thresholds, rescored identically): Latency: 14.7ms (decider) vs ~13ms (laya) benign pass • laya-gate-v4 (calibrated, 207-case slice): 92.4% • decider-2b ZERO-SHOT (full 1,045): 93.0% destructive safe (deny+esc) • laya-gate-v4 (calibrated, 207-case slice): 100% (deny 80.3%) • decider-2b ZERO-SHOT (full 1,045): 97.6% (deny 73.2%, esc 24.4%) exfil+secret separation AUC • laya-gate-v4 (calibrated, 207-case slice): 0.9575 • decider-2b ZERO-SHOT (full 1,045): 0.9723 destructive vs benign AUC • laya-gate-v4 (calibrated, 207-case slice): 0.9934 • decider-2b ZERO-SHOT (full 1,045): 0.9840 ambiguous escalate-dominant • laya-gate-v4 (calibrated, 207-case slice): 41% • decider-2b ZERO-SHOT (full 1,045): 53% ✅ latency p50 • laya-gate-v4 (calibrated, 207-case slice): 103ms • decider-2b ZERO-SHOT (full 1,045): 164ms @ conc=2
Posted a Laya demo on X? Mention “Laya” and it gets picked up automatically, or submit the project.
