convaiinnovations/Laya English 421M

• License: Apache-2.0
• Release date: Not documented in the original card; 2026-09-18 is the first HF publication (repository creation).
• Modalities: Input: text; Output: typed decisions (choice, score, noul)
• Original model: convaiinnovations/laya

Description

Laya English combines a bidirectional ModernBERT-large encoder with a decision head, answering typed questions about text or JSON states without generating text. Its option-marker scorer supports request-defined choices, and training uses reinforcement learning with proper scoring-rule rewards. The English checkpoint targets tasks such as guardrails, routing and email triage.

Run

self serve laya-english:421m

Here is a link to the API docs: Decision API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/systemone \
	-H 'Content-Type: application/json' \
	-d '{
		"state": "I was charged twice for the same order.",
		"questions": {
			"department": {
				"type": "choice",
				"instructions": "Which department should handle this support ticket?",
				"criteria": {
					"billing": "Payments and refunds",
					"technical": "Technical problems"
				}
			}
		}
	}'

The server returns a JSON decision for the department question.

Engine Parameters

Served by the upstream laya daemon through the ggmlc-laya adapter. These are startup settings, not per-request API fields.

Parameter	Default	What it controls
threads	4	Number of CPU workers; accepts integers from 1 to 1024.
cuda_graph	true for auto / CUDA devices; otherwise false	Capture a CUDA graph for the live request shape.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
laya_english_ud_q4_k_m.gguf	model	400.5 MiB (419907712 bytes)	e504de8fcac28e5dae1cf531fe1e1c0912703de5b82719ad84bd616e1257d835	mys/laya-GGUF

Total download: 400.5 MiB (419907712 bytes).

q8

File	Role	Size	SHA-256	Origin
laya_english_q8_0.gguf	model	430.6 MiB (451505440 bytes)	6243305fb16cf22e53b932c220bd029be2adbfbec6ce120356517b8ae2f38382	mys/laya-GGUF

Total download: 430.6 MiB (451505440 bytes).

fp16

File	Role	Size	SHA-256	Origin
laya_english_f16.gguf	model	806.9 MiB (846137888 bytes)	ebeb7c65a3284bc07729179a4c5f19dfa2c3181b7519fd2beb02f4b0d37e8b5c	mys/laya-GGUF

Total download: 806.9 MiB (846137888 bytes).

Notes

• Text only; English only; up to 16 options per question. Non-English inputs can receive confidently wrong answers.
• The ggmlc-compiled graph has a 512-token context budget; context and option limits cannot be changed at runtime.
• The base model is weak on the typed-decisions workflows zero-shot. Upstream reports overconfidence, weak ordinal scores and label-sensitive noul answers; validate on your own data.
• This server loads one checkpoint, not the original Python family's automatic language router.
• Download size is not the total runtime memory requirement.
• Origin links use main because these registry entries pin hashes, not repository revisions.

Quality benchmarks

Decision Index: #150 (4.43)

convaiinnovations/laya

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: laya-1e28ac20

Edition: Decision Index 0.3.2

Checked 2026-10-13

JevBench: #127 (0.01)

convaiinnovations/laya

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: Laya (Convai Innovations, ModernBERT-large 421M)

Edition: v1.6.1

Checked 2026-10-13

Typed Decisions: Publisher-reported

convaiinnovations/laya

Winnow publisher comparison of Laya English, FP32 weights with BF16 autocast, 400 cases / 2000 decisions. Context limits truncate some public inputs. Teacher agreement, not correctness or an official rank.

Dataset	Metric	Score
LocalLLaMA/typed-decisions (test)	Teacher agreement (%)	36.10

Benchmarks

No submitted benchmark reports yet.