115 builds · page 1 of 3
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！ https://t.co/31KGUiNunb
X postTools & apps
★ PickSeven open source Jev models tested in 21 minutes. Laya gets the last 3:41 0% - 0:00 - intro 17% - 3:46 - JevBench, the thing they all get scored on 33% - 7:02 - Nimble 51% - 11:03 - Decider 62% - 13:20 - OpenJev 83% - 17:50 - Laya The order is the story. Laya did 2M views this
X postLLM routing & guardrails
1,000 synthetic emails. Six folders. One local 322M Laya model running on a MacBook Air M3. ⚡ 28.6s without screen recording 🎬 97.3s while recording 🎯 65.1% accuracy against reference labels 💾 0 MB swap Everything runs locally. No cloud inference. And I’m showing the
X postSupport & triage
Beating Laya (https://t.co/rTxWDMS8KG) in tic tac toe : ) Using this for Unity https://t.co/XEWtg1ZaEd by @asus4
X postTools & apps
ScreenQuest: an autonomous screenshot-to-action AI agent for https://t.co/RHiWnwGPEf, running locally on an M3 Max (48 GB). Qwen3.5-4B / MLX + Laya / Core ML + Apple Vision OCR. No cloud inference. Code: https://t.co/gFP4sjd3VU
X postTools & apps
The open-source local version of Laya crushes Jev in response speed! On the left is Laya, a local 421M open-source decision model. On the right is Jev 1.13.0, running through a cloud API. Same Snake game, same typed decisions, and the same 30-second free run: - Laya: Score
X postTools & apps
I had a stupid idea: What if, instead of asking an #AI to talk, I made it fight me? 🥊 So I built World Summit Tournament. 21 fictional world-leader-inspired fighters, signature moves, combos, replays — and the opponent can use Laya, a small open-weight decision model in the
X postTools & apps
Two AIs. One game. I let @typesafeai 's Jev and @brainFnCl's Laya play the same arena survival game where every enemy’s decision is made live by the models. How it works: the game turns each moment into one sentence and asks one typed question. Back come probabilities, not
I connected Pac-Man to Laya-MLX 👻 Every move is a real typed decision running locally on Apple Silicon—no cloud API, no text generation, 0 output tokens. The live view shows action probabilities, latency, tactical planning, and safety interventions. { "steps": 351,
X postTools & apps
I made Laya (Jev-like model) play Breakout, Every paddle move is a live decision from an open 322M System One model (convaiinnovations/laya) running locally on Windows laptop, ~70 decisions/sec on an RTX 5080, zero cloud, $0. Code 👉 https://t.co/8v0dnXCnQ1
i made Jev and Laya fight 1v1 each other in Doom 1v1 deathmatch, monsters in between, first to kill the other opponent 5 times wins! @typesafeai Jev's calls were sharper, it needed a third of the corrections. @mizorewww Laya just decided twice as often. final results
X postTools & apps
If you’re wondering how fast Laya is compared to Jev: my maze demo measured 19ms median responses from local Laya on a base M5, versus roughly 300ms from Jev’s cloud API.
Tetris bench: Jev vs. Laya-mlx I run the Laya model on my MacBook. It's true that the model is really fast (~84ms), but it's also much more stupid. Losing to Jev 3 out of 3 rounds.
i made Jev and Laya play the chrome dinosaur game against each other except they’re actually controlling the dinosaurs. Laya runs locally on my Mac, Jev runs over an API. same track, same physics, completely different latency. they make their own decisions, dodge obstacles,
X postTools & apps
This is NOT Jev. Open source. Runs on your laptop. Decides in ~27 ms, about 200× faster than waiting on a hosted LLM. Here it is playing Tetris by itself 👇 https://t.co/eq4gP53o2A
X postTools & apps
The problem isn't the models, it's how you define the problem space, design the movements / decisions and optimize for it (batching e.g.) Jev wins for a network-based model any day. Local models tied - Laya-mlx and my custom build Jax. Look at them go.
X postTools & apps
Jev vs locally running Laya-MLX and Kev-4B I built my own snake benchmark. Jev @typesafeai is running via API. The other models are small alternatives running on very little RAM on my Macbook! Jev seems to deliver the best quality no doubt! After running it for a while 0
X postTools & apps
Laya beat Jev 43 to 1 with the wifi turned off 86.4 decisions per second against 3.2. 1,281 moves against 47 left side is Laya, an open weights decision model, 421 million parameters, running locally off a 1GB footprint. right side is Jev 1.13.0 over an API. both got the
Thanks for the model, we were able to port Laya to coreml with 99.5% of the ops on ANE + benchmarked too. it is now blazing fast with 3.7 ms per decision on an M5 Pro. Release: https://t.co/9xerYIHt9y Models: https://t.co/kPOaMQI4RH
Local Laya moggs Jev at @grok 4.7-built Tetris 🧩 An open-weights System One model called Laya, beat cloud-based Jev at playing Tetris by making decisions 11 times faster, running locally on a 16GB MacBook Air! Run AI models locally -> https://t.co/RbcCOIgVkj
A Myanmar-language Laya checkpoint trained on Davlan/sib200, with a Gradio demo space from the same author.
GitHubMultilingual
An Arabic Laya checkpoint trained on AmazonScience/massive and facebook/xnli, with a companion laya-ara-rag ranking model from the same author.
GitHubMultilingual
A Laya checkpoint trained on PolyAI's Banking77, the 77-intent banking support dataset, tagged for customer-support intent classification.
GitHubSupport & triage
AXERA-TECH's conversion of all three Laya checkpoints for its edge AI chips, reported at latency under 70 ms per decision on AX650/AX8850.
GitHubTools & apps
Google's LiteRT community conversion of the multilingual checkpoint (tflite), tagged for Android text classification. An English LiteRT build is published alongside it.
GitHubMultilingual
Fan-out SUMs, = NULL, UPDATE without WHERE, injection remnants and more, each with intent and schema context. Four typed questions per statement, identical payloads to both models.
GitHubSecurity & fraud
Synthetic Chinese workplace-chat scenarios with frozen inputs, prompts and labels. Reports classification quality, mis-generated tasks and latency: Jev 253 ms vs Laya 151 ms.
GitHubMultilingual
20 real-world Chinese requests × 3 checkpoints, measuring accuracy, latency and whether confidence knows when it is wrong. Includes a Windows/Python 3.14 segfault workaround.
Automotive, embedded, enterprise software and product decisions, three trials each, with questions, expected answers, raw responses and a static HTML report.
GitHubMultilingual
Reproducible zero-shot benchmark on informal Moroccan Darija reviews in Arabic script and Arabizi, positive/neutral/negative. Deterministic seed-42 split, no training.
Postgres FTS, pgvector and RRF first stage; Jev, Voyage, GPT-4.1-mini, GPT-5.6-luna and open-weight Laya as rerankers on two Finnish legal corpora. Recall, MRR, nDCG, latency.
Seven zero-shot extraction and classification families demoed, benchmarked and cross-compared, with sizes, licences and costs in one table.
GitHubDocuments & data
Frozen protocol, raw predictions, calibration and latency for Jev 1.13, PrismNLI-0.4B and Laya. Laya ties Jev on dair-ai/emotion and trails by 10–33 points elsewhere.
GitHubDocuments & data
Six classifiers over SST-2, AG News and more, reporting accuracy, macro-F1, ECE, p50/p95 latency, throughput and cost side by side. Laya runs locally on MPS or CPU.
GitHubDocuments & data
Matched-budget fine-tunes on Banking77 and five more public workflows (prompt injection, SMS spam, emotion, request routing), plus two Aegis chat-moderation fine-tunes.
GitHubModeration & safety
Independent Jev vs Laya measurements plus a cascade: Laya decides first, high confidence executes, low confidence escalates to Jev. At a 0.60 threshold it matched Jev's 78%.
GitHubLLM routing & guardrails
MCP server plus HTTP sidecar: token-budget preflight that reports what would be cut, a persisted calibration store, structured errors, and a noul fix (P(true) 0.5 → 1.0).
Anthony Alcaraz's talk demo routes typed decisions through Jev and shows Laya's answers beside them. On the TSMC run Laya rejected every true claim; that's the point.
GitHubDocuments & data
Local 15x15 Gomoku on Apple Silicon. Rules in Python, board in the browser, every AI move one real Laya choice over up to six candidates, probabilities shown. No Gomoku training.
A port of the jev-ultrafast browser agent that decides with laya-mlx instead of the hosted Jev API. DOM snapshot, executor, safety checks and inspector are the originals.
GitHubLLM routing & guardrails
