convaiinnovations/laya: Multilingual Non‑Autoregressive Decision Model
A 100+ language, non‑autoregressive text‑classification model that returns calibrated probabilities in a single forward pass, with fast inference and optional server integrations. 4,961 Hugging Face likes in 15 days since convaiinnovations/laya was created (340.0/day). 1 source across 1 channel. Automatically detecte
49Fact cited ·Analysis our interpretation ·Estimate forward-looking · [n] links to the source
4 claims drafted by openai/gpt-oss-20b were kept because each quote was found verbatim in its source; 3 unverifiable claims were discarded. Statements marked AI await human review.
What happened
- Fact Model is non‑autoregressive and returns calibrated probabilities in a single forward pass (~33 ms).[1]AI
Laya Multilingual, non-autoregressive System 1 decision model. Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.
- Fact It never generates text, eliminating hallucination.[1]AI
It never generates text, so there is nothing to parse and nothing to hallucinate.
- Fact Optional extras include HTTP server, MCP server, LangChain/LangGraph, ONNX Runtime, and TileLang GPU fast path.[1]AI
Optional extras: laya[serve] (HTTP server), laya[mcp] (MCP server), laya[langchain] (LangChain and LangGraph), laya[onnx] (ONNX Runtime), laya[fast] (TileLang GPU fast path).
- Fact convaiinnovations/laya model card is listed among trending models on the Hugging Face Hub.[1]
- Fact That is 4,961 likes in 15 days, an average of 340.0 per day.[1]
Why it matters
- Analysis Use the laya package for inference, fine‑tune on domain data, or deploy via the optional server extras for scalable applications.AI
- Analysis Automatically detected from public activity data. Not yet reviewed by an analyst.
Should I care?
What changed?
- Fact Model is non‑autoregressive and returns calibrated probabilities in a single forward pass (~33 ms).[1]AI
Laya Multilingual, non-autoregressive System 1 decision model. Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.
- Fact It never generates text, eliminating hallucination.[1]AI
It never generates text, so there is nothing to parse and nothing to hallucinate.
Who is affected?
- Analysis Developers building decision‑making or classification systems that require calibrated probabilities across many languages.AI
What can developers do with it?
- Analysis Use the laya package for inference, fine‑tune on domain data, or deploy via the optional server extras for scalable applications.AI
Is it production-ready?
Insufficient evidence
Analysis Not assessed.
What would make it important?
Not yet assessed.
What could prevent adoption?
Not yet assessed.
Technical change
- Fact Model is non‑autoregressive and returns calibrated probabilities in a single forward pass (~33 ms).[1]AI
Laya Multilingual, non-autoregressive System 1 decision model. Give it a state (text, email, ticket, or JSON) and typed questions; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.
- Fact It never generates text, eliminating hallucination.[1]AI
It never generates text, so there is nothing to parse and nothing to hallucinate.
- Fact Optional extras include HTTP server, MCP server, LangChain/LangGraph, ONNX Runtime, and TileLang GPU fast path.[1]AI
Optional extras: laya[serve] (HTTP server), laya[mcp] (MCP server), laya[langchain] (LangChain and LangGraph), laya[onnx] (ONNX Runtime), laya[fast] (TileLang GPU fast path).
Architecture
No architecture has been documented for this signal.
Developer impact
Developer implications have not been assessed yet.
Difficulty has not been estimated yet.
Adoption & momentum
No adoption indicators recorded.
Timeline
First seen by the Astrava pipeline[1]
Limitations & risks
- Fact Accuracy on very long documents (>4,000 tokens) varies and should be validated.[1]AI
In the table above, 16 to 18 of 20 requests were answered correctly with up to about 4,000 tokens of text before them; beyond that results vary (8 to 17 of 20), so check long-document accuracy on your own data.
- Analysis Generated from activity metrics only; verify the linked sources before acting on it.
Competing approaches
No competing approaches documented yet.
Who is building on it
No independent projects or organisations are documented as building on this yet.
Evidence & sources
Primary source (1)
Published by the people who built it: release notes, repositories, papers, official docs.
- [1]convaiinnovations/laya model card(opens in a new tab)
text-classification · transformers · safetensors · laya · system-one · calibrated-decisions · rlcd
Secondary source (0)
Independent reporting, third-party benchmarks, integrations by other projects.
None recorded.
Community signal (0)
Discussion threads and social posts. Evidence of attention, not of correctness.
None recorded.
How the score of 49 was calculated
| Factor | Share of maximum | Points |
|---|---|---|
| NoveltyRubric level 2/4 (Moderate). Heuristic (automated): first published 15 days ago. | 7.5/15 | |
| Technical significanceRubric level 3/4 (High). AI-assisted (openai/gpt-oss-20b), pending human review: Provides calibrated probability outputs without text generation, enabling trustworthy decision systems at scale. Evidence: “Trained with reinforcement learning against strictly proper scoring rules (RLCD), so reporting honest probabilities is the only way to maximise reward.” | 11.3/15 | |
| Developer relevanceRubric level 3/4 (High). AI-assisted (openai/gpt-oss-20b), pending human review: Offers fast, language‑agnostic inference with optional server integrations, suitable for production use. Evidence: “It returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.” | 11.3/15 | |
| Adoption velocityIndependent integrations: unknown (0 pts). Hugging Face likes: 340.0 per day since creation; full marks at 1000/day, log-scaled (4.2/5). | 4.2/10 | |
| Community momentumFewer than three periods of activity data, or no baseline activity to compare against. | 0/10 | |
| Source credibilityPrimary source present; mean source-tier weight 1.00 (primary 1.0, secondary 0.7, community 0.35). | 10/10 | |
| Cross-source confirmationNo independent publisher has confirmed this yet. | 0/10 | |
| Evidence strengthMet 2/4: primary source ✓, runnable artifact (code, weights or package) ✓, reproducible or independent benchmark ?, technical documentation or paper ✗. 1 unknown (scored 0). | 5/10 | |
| Time to impactNo time-to-impact estimate yet. | 0/5 | |
| Hype penaltyNo hype indicators detected. | 0 | |
| Breakout Score | 49 |
Low confidence (0.48) — 60% of measured factors have data. 0 independent non-community publishers. At least one primary source. Evidence strength 5/10. Capped at Low: no independent, non-community source yet.
An analytical ranking derived from observable signals using model breakout-v1.2. It is not a prediction and not a measure of quality. How it works