Laya alternatives
Jev was released in September 2026, and within weeks several open models with the same typed-question interface appeared. Laya is a 322M to 421M encoder. Most alternatives are open LLMs, mostly Qwen, fine-tuned at 0.5B to 35B parameters to score options without generating text. They trade speed for accuracy.
472 projects
sysone-cpu
TayHCode
Runs Laya typed-decision checkpoints on CPU with PyTorch or optional ONNX Runtime
DecideLint
Eileenes
Tests, calibrates and sets abstention thresholds for typed-decision models including Laya
verdict
ashiksharonm
Fine-tunes Laya for multi-check LLM guardrails and evaluates it with open benchmarks
jey_pro_v2
buibui2026
Builds and evaluates a Laya fine-tuning baseline for legal RAG judging
system-one-harness
mev0lent
Compares Laya and other decision models as gates for HTTP response triage
laya-benchmark-poc
BanHun28
Independent benchmark of Laya 0.3.23 on intent, typed-decision and support-ticket tasks
laya-pharma-surveillance
ericmjl
Uses runnable notebooks to evaluate Laya on pharmaceutical surveillance workloads
benchmark-llm-iot-laya-jev
alinembs
Compares local Laya and hosted Jev decisions for controlling MQTT-connected lights and a fan
Laya-LLM-Fine-tuning-and-benchmark-evaluation
sudhi-hub
Fine-tunes Laya for typed decision tasks and evaluates it with benchmark datasets
decision-model-arena
dante01yoon
Compares four decision models, including local Laya, across games and image-based decisions
d1-ane
nathanrchn
Converts and benchmarks decision models on Apple Neural Engine, with a Tetris demo
laya-comparison
felipegoodman
Compares Laya with a lightweight local language model
decision_api_benchmark
guming3d
Benchmarks decision engines, including Laya, on model routing accuracy, latency and cost
layatrade2
fearlessfly786-png
Fine-tunes and evaluates Laya System One on BTCUSDT trading decisions
layatrade3
fearlessfly786-png
Trains and evaluates a two-head Laya trading-decision model
rlcd-tool-selection
VectorSophie
Compares Laya and Jev tool selection with native and decomposed LLM tool-calling approaches
The models
Sizes and licences are from each model's Hugging Face card.
| model | maker | sizes | licence |
|---|---|---|---|
| Kev | Jared Palmer | 0.5B to 27B (benchmarked: 0.8B, 4B, 9B) | Apache 2.0 |
| Decider | Mapika | 0.8B to 35B, plus a 2B vision model | Apache 2.0 |
| Clef, Clef-flash | Cloudflare | 27B, 9B | Apache 2.0 |
| GLiNER2.5-Decide | Fastino | 340M English, 287M multilingual, 1B | Apache 2.0 |
| Julia-1 | Supersonic Labs | 144M | Apache 2.0 |
| Jebadiah | Frontier Infra | 4B, 9B, 27B | Apache 2.0 |
| pplx-decider | Perplexity | ~26B ("27b" in the name) | Apache 2.0 |
- Kev runs a frozen Qwen3.5 base prefill-only, with a LoRA adapter and a pointer head over the options. The card says states of up to 65,536 tokens are served and 8,192 are validated.
- Mapika Decider calls itself an open reproduction of the System One model class. Its training code and data registry are on GitHub.
- Clef and Clef-flash read text, JSON, images and video, per their cards. Cloudflare's announcement says they are Jev-API compatible and also served on Workers AI.
- GLiNER2.5-Decide is a classification-only member of the GLiNER2.5 family: any label set at call time, single- or multi-label.
- Julia-1 is, like Laya, an encoder: it starts from mmBERT-small. See ModernBERT and mmBERT in Laya.
- Jebadiah ships a standalone server that answers
POST /v1/systemoneon one CUDA GPU or an Apple Silicon Mac. - pplx-decider v1.1 uses a separate decision readout instead of a normal language-model head, so its card asks you to use the included inference code.
The hosted reference is Jev from TypeSafe. Its weights are closed and it is billed per input token. The differences from Laya are in Laya vs Jev.
What independent benchmarks found
akitaonrails (October 10, 2026) used synthetic, AI-assisted cases in English and Brazilian Portuguese without independent human labels. Chance was about 31%. Local models ran on a Ryzen AI Max+ 395. Compact-set accuracy and median latency:
| model | accuracy | latency |
|---|---|---|
| Jev (hosted) | 97.6% | 255 ms |
| Clef (hosted) | 95.8% | 302 ms |
| Perplexity Decider 1.1 | 99.4% | 1,547 ms |
| Jebadiah 4B v2 | 90.5% | 205 ms |
| Kev 9B | 88.1% | 561 ms |
| Kev 4B | 87.5% | 246 ms |
| Mapika Decider | 87.5% | 186 ms |
| Clef-flash (local) | 87.5% | 618 ms |
| Kev 0.8B | 63.1% | 86 ms |
| GLiNER2.5-Decide (English only) | 52.4% | 208 ms |
| Laya (English) | 42.9% | 41 ms |
| Julia-1 | 31.5% | 12 ms (GPU) |
The author's conclusion: if Jev-like accuracy is a requirement, Laya, Julia-1 and the 0.6B Decision 3.0 Edge are not usable out of the box. He also notes that judging these small encoders by the accuracy column alone misreads them: they are built to be fine-tuned and run at the edge. Fine-tuning was not tested.
Arize (October 2026) ran a bracket of hallucination detection on RAGTruth, LLM-evaluator suites and agent routing. Balanced accuracy on RAGTruth was 85.5% for Jev, 85.2% for Clef 27B, 76.8% for Kev 9B and 59.8% for laya-multilingual. Laya was the fastest model at 0.04 s p50 on an M4 Max, and had the highest calibration error (0.330).
systemonemodels.ai lists 29 open Jev alternatives with the accuracy each publisher reports. Those figures come from different test sets, so they cannot be compared with each other.
How to choose
- Speed, CPU or on-device: the encoders. Laya and Julia-1 answer in tens of milliseconds, and Laya has the most ports, from Mac to the browser. Plan to fine-tune or to keep a fallback for hard cases.
- Accuracy without training: the 4B and larger models. They need a GPU or an Apple Silicon Mac with enough memory, and they are several times slower.
- No servers: Jev, Clef on Workers AI, or a hosted Laya API.
Whichever you pick, run a few hundred of your own labelled examples through two or three candidates. The rankings above come from small or synthetic sets, and their authors say so.
More ways to use Laya