Catalog, page 3

githubMJ

@githubMJ

↗

Java port of Laya with choice/score/noul, a script-detecting router that picks the checkpoint, HuggingFace Hub fetching and bit-level parity with the Python ONNX output.

li-ming1

github.com

↗

Tokenizer, mmBERT-base encoder and decision head in one binary for the 322M multilingual checkpoint. No Python, torch or BLAS; forward pass checked against a reference.

Mahmoud Shehata

@b0xtch

↗

Loads original Laya safetensors with Candle and runs choice/score/noul with fused kernels, tiled attention and GEMM dispatch tuned for Apple Silicon. Optional f16 on Metal.

Paperinik

@aovestdipaperino

↗

No Python, no torch, no ONNX export. Loads the ungated safetensors directly and runs any of the three checkpoints on CPU, Metal or CUDA.

Lars Karlslund

@lkarlslund

↗

ggml-based C++ runtime for all three checkpoints. Native tokenization, decision heads and JSON output; optimized FP16/BF16 CUDA on RTX with FP32 accumulation, plus Vulkan.

Receptron

@receptron

↗

npm package that runs Laya with ONNX Runtime, no Python or PyTorch at runtime. Same request shape as TypeSafe's system_one API; output matches Python to four decimals.

mizorewww

@mizorewww

↗

Validated Core ML ports of all Laya checkpoints. One short multilingual decision in 4.98 ms P50 on M3 Max with ANE FP16, at 2.78x better energy per decision than MLX FP16.

Nandakishor m

@Nandakishorm1

Laya surpasses jev when it's fine tuned in almost all cases..

Nandakishor m

@Nandakishorm1

How I built Laya

Nandakishor m

@Nandakishorm1

Asking everyone to fine tune https://t.co/mwPvWGPY2n 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

Abhijay

@abhijay_cloaked

There’s something cute about giving a probabilistic classifier a freshman probability exam. So I gave @typesafeai Jev, SemIf, and Laya ~1000 adapted questions on probability and countability from Berkeley exams Jev: 83.7% SemIf: 61.6% Laya: 31.2% A little stress test for the

Sai Dutta Abhishek Dash

@AbhishekDash69

I benchmarked Jev 1.13.0 vs open-weights Laya on 751 identical questions. Jev sweeps triage/guardrails/moderation, goes 1.000 on 5-language intent. Laya wins agnews + mnli at $0. Full data:

iamsaturday

@my_small_room

Tested Jev vs Laya (local, MLX) on 40 support tickets, EN + Thai. Jev 90–100%. Laya zero-shot 30–80%. Laya's faster, but not close on accuracy yet.

Madhav Sharma

@MadhavSz

Laya shipped right after Jev claiming to be faster for AI agent decisions. curious if it was accurate too I built a POC same Google ADK triage agent, same 30 labeled tickets, only the engine swapped. Jev: 422ms, 53% accuracy. Laya: 152ms 10% accuracy Faster not more accurate

FrontierMind

@Raju_M_A

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.

Aaron Edell

@aaronedell

Benchmarked @mizorewww’s Laya-MLX against @FeatherlessAI Simple Jev on my M4 Pro inside @OpenClaw: Laya: 85% @ 27.8ms Jev/Qwen3-4B: 90% @ 1,167.6ms The stack is rules → Laya → Jev → @ollama 27B → cloud. 14/20 decisions resolved before Jev. @typesafeai was right about System

Andrea Bruno

@baronunread

I've added Laya and https://t.co/flwFG40TX4 (@michael_chomsky) support to leanest, Jev is good but having other options to try out seems nice. Also new og image (I hate og image caching so bad) https://t.co/OgVzZg8CVE

VeidtKV

@VeidtKV

I ported laya to my specific hardware (6900XT) with the help of my hermes agent. Check it out here

Anton Lobach

@AntLobach

Wow, Laya (https://t.co/PhmZAwTZeX) inference in Jolt (Clojure running on Chez Scheme)

Akihito Koriyama

@koriym

Built a hypermedia browser on Laya-MLX — a small model that, like Jev, doesn’t generate text; it just returns a probability for each option you give it. The next link is chosen from HAL’s _links and what ALPS declares those links mean. No routes defined up front. cc/ @mamund

//s̶̀pò̶of.sh

@spoofydude

I'm publishing this workflow! a 3-model fully local pipeline (prompt enhancement + Laya decision-making) that generates an image in ~37s (fast af) Should run great on any RTX 3080+ https://t.co/i11GcPJbIm

pythongiant

@pyth0ngiant

Built laya-drift to visualise and monitor agentic-drift today. It continuously measures how far an OpenCode session has drifted from its original plan using Laya + Jensen-Shannon divergence https://t.co/eB9DNlQksp

Astrid

@just_some_dev

I made a postgres plugin for the Laya decision models, using @Kelwing's Rust framework for running them

Leonardo Stenico

@50bbx

I forked Needle (the Chrome extension) to use Laya instead of Jev. Fully local AI browser search. Try it out here: https://t.co/SYB7HQfya9 It is not a semantic search in strict terms, it will match your search to semantically similar results.

Shantanu Goel

@shantanugoel

Update: I trained Laya non-step for close to 24 hours. No where close to jev 0-shot :|

TTUX | تی‌تاکس (AmirTaha سابق)

@TTUX_tech

جالب شد بازی اسنیک. ورودی لوکیشن میوه و هد و بوردر هارو میگیره و با مدل اوپن سورس #Laya به صورت لوکال که شبیه #Jev هست دستور میگیره که لوکیشن بعدی کجا بره بالا پایین چپ راست. تجربه جالبی بود اینم لینک پروژه تو گیت هاب ممنون میشم با 🌟 حمایت بکنین: https://t.co/dcQcs6WljY

Mansour Raad

@mraad

The engineer and the trainee. Adaptive MPC knows the physics. Laya, a small decision model, learned from examples. Alone, Laya lands 0/90. With MPC vetoing about one in three of its moves, it gets 90/90, through a 60% engine loss. Fly it in your browser:

thaitea

huggingface.co

↗

A vision variant of Laya on Hugging Face: the same choice/score/noul question shape, with an image as the state.

Mattepiu

huggingface.co

↗

ONNX export of the Laya checkpoint so it runs anywhere onnxruntime does, including browsers and edge boxes. An fp16 variant followed a day later.

mys

huggingface.co

↗

GGUF conversions of both the English and multilingual Laya checkpoints for llama.cpp-style local runtimes.

ShaunSpark

huggingface.co

↗

A browser-use agent that picks the next action with Laya, trained and evaluated on the Mind2Web web-task benchmark. Weights on Hugging Face.

雨夹雪

@mizorewww

↗

MLX port of Laya with performance work. Under 1 GB of memory on-device; the demo plays Snake at 60 decisions per second on an M3 Max.

Nandakishor’s dev.to write-up on building decision models a year before Jev, RLCD training against proper scoring rules, and why Laya is open.

Gradio demo of Laya on Hugging Face: email triage, phishing, guardrails, ticket routing, RAG filtering and multilingual routing, no install.

Nandakishor Mukkunnoth

@NandhaKishorM

↗

The model itself. Typed choice/score/noul questions over any state in one forward pass, 100+ languages, Apache 2.0 weights.