Laya 拼音输入法（macOS）用的候选词排序模型，基于 convaiinnovations/laya（revision 7b928d82）微调。 prior 对于词来说是词频，对短语来说是词格路径得分。模型看不到用户输入的拼音。 训练样本由输入法自己的引擎（backend/，commit 9d87301）录制，每条样本都对应一次真实的模型调用：上文、按引擎 - 从基座 Laya 开始，除 embedding 以外的全部参数参与训练。 - 损失：对每次调用的候选做 listwise cross-entropy，得分为 logit + betakind prior，每种调用类型各学一个 beta；每个 epoch 打乱候选顺序。 - AdamW，LR 3e-5，weight decay 0.01，bf16 autocast，gradient clipping 1.0。 - 学习率：1,000 step warmup，之后保持不变，最后 20% 线性衰减到 0。 - 166,432 step，batch 100（1,664 万条，约 1.15 轮），单张 RTX 5090 约 6 小时。 bootstrap 区间。对照为未公开的上一个版本 w28。 - 依赖本引擎： 模型针对本引擎的候选列表和 prior 微调，不是通用语言模型，不能单独当输入法用。 权重为 Apache 2.0，是 convaiinnovations/laya（Apache 2.0）的衍生作品。inference/、scripts/ 和 eval/ 中的代码为 MIT，与输入法仓库相同。
SAVRN Model Hub
Open-Weight Models
An open-weight model is an AI model whose trained weights are published for anyone to download. The weights are what the model learned in training. With a copy of them you can run the model on hardware you control and train it further on your own data.
Open weights are not the same as open source. Many publishers release the weights without the training data or code, and the license sets what you may do with the model. This library puts each model's full card, architecture, files, license and published evaluations on one page.
11,442Models
3,247Datasets
285Papers
6,976Publishers
12,091Sourced relationships
Updated 2026-10-19 · How the library is built
11,442 models, sorted by most downloaded.
This checkpoint fine-tunes convaiinnovations/laya-multilingual for native choice, score, and noul decisions in Portuguese and Spanish. It keeps the original 322M-parameter mmBERT architecture. It adds no inference component and does not generate text. It returns typed answers and probabilities in one forward pass. This is a text model. Inference takes a textual state plus typed questions. The second training stage used text decisions derived from public speech corpora, but this checkpoint does not accept audio by itself. The separate audio projector is not included. The official Laya SDK defines these primitives as follows: - choice: selects one key from a runtime-defined criteria object.…
Fine-tune of convaiinnovations/laya (subfolder="multilingual") for Thai typed decisions. The model does not write text. One forward pass stamps choice, score, and noul (yes/no) answers and returns a probability for each. The weights in this repo are epoch02. The config is the calibrated one: same weights, softer confidence. A single temperature does not change the chosen label. This release is published by the Hugging Face organization that owns the repo. No individual author is named. The model is a fine-tune, not a new base: Comparison numbers for iapp/OpenThai-SystemOne in the speed table were measured locally from those weights. iApp is not an author of this fine-tune. Route a Thai…
Classifies GitHub issues written in any language as bug, feature, question or docs. A fine-tune of Laya multilingual (mmBERT-base) used by the laya-triage GitHub Action for non-English issues, next to the English model laya-triage-en. The same 500 NLBSE'23 validation issues, machine-translated with NLLB-200 into 13 languages. Accuracy (±3 points per language): laya-triage and Jev are within noise of each other across languages; both are far ahead of the untuned base. Translations can flatter a model trained on translations, so we also checked real issues: on 367 non-English issues opened in 2026 (never seen, written by people, not translated) accuracy went from 47.1% to 65.7%. Use it…
Non-autoregressive System 1 decision model, fine-tuned on the typed-decisions workflows: agent-trace observability, customer service, invoice processing and security incidents. Part of the Laya family. 400 test cases, 2,000 decisions, measured on the official test split. +3.9 points over Jev's published 0.727, above the 0.735 teacher ceiling, with 2.4x better Brier and 1.6x better score MAE. Jev figures are third-party published, not measured here — there is no TypeSafe API access in this project, and sample sizes and prompts differ. Treat the comparison as indicative. Router will not select this checkpoint automatically unless you construct it with autotaskdetection=True — it is…
Laya typed decisions on Apple Silicon, using CPU + GPU. This is a portable Core ML bundle for laya-coreml, converted from convaiinnovations/laya-typed-decisions. It outputs choice, score, and noul probabilities with zero generated tokens. Inference needs no PyTorch, Transformers, MLX, remote code, or cloud API. Apple Silicon, macOS 15+, Python 3.11–3.13. Tested on M3 Max / macOS 27.2. To download explicitly and then run entirely offline: Use laya.load("aac6fef/laya-typed-decisions-coreml", localfilesonly=True) for a cached snapshot or pass a local directory. Use revision=" " to pin a remote revision. This FP16 export retains the original model architecture and decision schema. The…
A 10,284,480-parameter LLaMA-style text model, trained from scratch. This is a verified first checkpoint for the LDT-10M request (model-requests #12, DedeProGames) — real weights, real training, but undertrained (see the honest status below). It is not a quality release yet; the card states that plainly. Standard LLaMA block, no sliding window, no GQA: The parameter count is the learnable total: the raw safetensors sum is 14,216,640, which double-counts the tied embedding (tok.weight 12288×320 = 3,932,160) that head.weight aliases. Tied, the true count is 10,284,480. - Trained from scratch (no base model). The model learned real context — val loss 4.602 is well below the 7.38 unigram floor…
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
The retained trained Legal head was replayed on four development fixtures. All 32 training and 8 tuning numerical inputs matched its retained historical manifests before an explicit source-binding compatibility fork. No fresh held-out accuracy claim is made for this package. This is an opt-in development artifact. It is not the default symbolic compiler, not a verified formalizer, and not a security repair policy. All qualification, admission and proof-authority flags remain false. No LLM/provider fallback runs. Input embeddings use thenlper/gte-small at commit 17e1f347d17fe144873b1201da91788898c639cd (384 dimensions). The embedding encoder weights are not included. Source inference…
This lane contains 7 distinct original checkpoint variants. Each checkpoint uses a content-addressed path and an asset manifest preserving its schema, decoder role, available format/profile binding, codec, source declarations, and companion corpus or embedding assets. Unknown bindings remain unknown. Registration and publication do not qualify numerical quality, source-only reconstruction, teacher use, proof authority, or serving. No retained checkpoint is activated for inference. The Legal 768D checkpoint is inherited initialization only; Codebase, Intent, and Security 768D lanes have no retained checkpoints. Codebase 384D envelopes retain their borrowed Security payload identity and are…
This lane contains 1 distinct original checkpoint variants. Each checkpoint uses a content-addressed path and an asset manifest preserving its schema, decoder role, available format/profile binding, codec, source declarations, and companion corpus or embedding assets. Unknown bindings remain unknown. Registration and publication do not qualify numerical quality, source-only reconstruction, teacher use, proof authority, or serving. No retained checkpoint is activated for inference. The Legal 768D checkpoint is inherited initialization only; Codebase, Intent, and Security 768D lanes have no retained checkpoints. Codebase 384D envelopes retain their borrowed Security payload identity and are…
This lane contains 1 distinct original checkpoint variants. Each checkpoint uses a content-addressed path and an asset manifest preserving its schema, decoder role, available format/profile binding, codec, source declarations, and companion corpus or embedding assets. Unknown bindings remain unknown. Registration and publication do not qualify numerical quality, source-only reconstruction, teacher use, proof authority, or serving. No retained checkpoint is activated for inference. The Legal 768D checkpoint is inherited initialization only; Codebase, Intent, and Security 768D lanes have no retained checkpoints. Codebase 384D envelopes retain their borrowed Security payload identity and are…
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Use the code below to get started with the model. Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019). - PEFT 0.21.0
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,511 tokens. The source MXFP4 checkpoint was explicitly dequantized to BF16 before scoring and pruning. The source-row selection hash is…
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…
This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqamain configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The released loader settings are preserved: train, Question plus shuffled choices, Explanation, selectionseed=1234, BF16, and no optimizer step. The samples are full length: no tokenizer maxlength, truncation, or padding. The longest input for this tokenizer is 1,632 tokens. The source checkpoint was loaded and pruned in BF16. The source-row selection hash is…
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
Embedding models for retrieval
Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher.
Collection · 4 entries
Models that fit on one accelerator
Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer.
Collection · 6 entries
Open-weight text models worth knowing
Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
Open-Weight Models Explained
What is an open-weight model?
An AI model whose trained weights are published for anyone to download, so it can be run, tested and fine-tuned on hardware the user controls.
Is an open-weight model the same as open source?
Not always. Open weights means the trained model can be downloaded. Open source usually also means the training code and data are available and the license allows broad reuse. Many open-weight models release the weights only.
Can I use an open-weight model commercially?
It depends on the license. Apache 2.0 and MIT allow commercial use. Other licenses limit it, for example to non-commercial use or below a set number of users. Every model page here shows its license.
How much memory does an open-weight model need?
About two bytes per parameter at 16-bit precision, so a 7-billion-parameter model needs roughly 14 GB for its weights, plus memory for the context it processes. Each model page lists its parameter count and the size of its files.
Related SAVRN Research
The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them.
SAVRN Index
What open models cost to run
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.