ModernBERT-large encoder (421M total) with a from-scratch decision head: option markers scored at their own MASK token, softmaxed per question. Answer schema is defined at request time, so new question types need no retraining. Three checkpoints routed automatically: English (512 ctx), multilingual mmBERT 322M (1024 ctx, 100+ languages), typed-decisions (1024 ctx). RLCD training. 33ms single-question answers on a T4.
AI-generated content marks
The provider reports that this model does not add embedded watermarks or provenance metadata to generated output.
- 420M
- 1k
- apache 2.0
- Sep 2026
Guides covering Laya
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | Q4_K_M | Q8_0 | BF16 |
|---|---|---|---|
| 4x H100 80GB (320GB) | fast 29480.0t/s | fast 16377.8t/s | fast 8773.8t/s |
| NVIDIA DGX Station 748GB | fast 17600.0t/s | fast 9777.8t/s | fast 5238.1t/s |
| 8x RTX 3090 rack (192GB) | fast 16477.1t/s | fast 9154.0t/s | fast 4903.9t/s |
| 4x RTX 5090 (128GB) | fast 15769.6t/s | fast 8760.9t/s | fast 4693.3t/s |
| AMD Instinct MI300X (192GB) | fast 11714.6t/s | fast 6508.1t/s | fast 3486.5t/s |
| 4x RTX 4090 (96GB) | fast 8870.4t/s | fast 4928.0t/s | fast 2640.0t/s |
| 2x RTX 5090 (64GB) | fast 7884.8t/s | fast 4380.4t/s | fast 2346.7t/s |
| 2x RTX 3090 (48GB) | fast 4119.3t/s | fast 2288.5t/s | fast 1226.0t/s |
| Single RTX 5090 (32GB) | fast 3942.4t/s | fast 2190.2t/s | fast 1173.3t/s |
| RTX PRO 6000 Blackwell (96GB) | fast 3942.4t/s | fast 2190.2t/s | fast 1173.3t/s |
| Mac Studio M4 Ultra 192GB | fast 2620.8t/s | fast 1456.0t/s | fast 780.0t/s |
| Mac Studio M4 Ultra 512GB | fast 2620.8t/s | fast 1456.0t/s | fast 780.0t/s |
| Single RTX 4090 (24GB) | fast 2217.6t/s | fast 1232.0t/s | fast 660.0t/s |
| MacBook Pro M5 Max 128GB | fast 1473.6t/s | fast 818.7t/s | fast 438.6t/s |
| Single GTX 1080 Ti (11GB) | fast 1064.8t/s | fast 591.6t/s | fast 316.9t/s |
| Dual EPYC 9004 + 768GB DDR5-4800 | fast 1013.8t/s | fast 563.2t/s | fast 301.7t/s |
| DGX Spark 128GB unified | fast 600.6t/s | fast 333.7t/s | fast 178.8t/s |
| Ryzen AI Max+ 395 128GB | fast 563.2t/s | fast 312.9t/s | fast 167.6t/s |
| Jetson AGX Orin 64GB | fast 450.6t/s | fast 250.3t/s | fast 134.1t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | fast 450.6t/s | fast 250.3t/s | fast 134.1t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | fast 337.9t/s | fast 187.7t/s | fast 100.6t/s |
| NVIDIA Jetson Orin NX 16GB | fast 225.3t/s | fast 125.2t/s | fast 67.1t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
Download options
Or run it in the cloud
No per-token API provider pricing tracked for Laya yet. For flagship list prices, see the calculator.
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.