Laya had reached 794 points and 184 comments on Hacker News by 13:30 UTC on September 19. Its GitHub repository, created the previous day, had also crossed 1,000 stars by the time of reporting. The number that needs closer attention is 0.952: Laya's English checkpoint reported 95.2% mean confidence while getting every Khmer example wrong in one published test. That failure appears in the project's own benchmark report, a useful piece of evidence that cuts against the usual launch-page polish.
ConvAI Innovations released Laya days after TypeSafe AI introduced Jev, a closed early-access model for fast, typed decisions. The Laya launch page presents the project as an open counterpart, with an Apache 2.0 Python package, downloadable weights and reproducible test files. That makes this release more than a second explanation of the same model category. Developers can inspect where its probabilities fail, rerun the harness and decide whether a narrow decision model belongs in their own request path.
A model that refuses to write
Laya accepts text or structured state, evaluates a set of questions and returns probabilities in one forward pass. Its three output primitives mirror Jev's contract: choice selects from named options, score estimates a position on an ordered scale, and noul returns a probability for a Boolean question. The repository documentation suggests support routing, prompt screening and email triage as possible jobs. It generates no prose, so an application does not have to extract an enum or repair malformed JSON after the call.
A fixed schema prevents an invented field, though it cannot prevent a wrong answer that fits the schema. Laya's own benchmark report makes that distinction concrete. On a held-out moderation set, its best checkpoint reached 0.530 accuracy and 0.400 macro F1. The same report says the demo's hand-picked moderation examples work while the broader test does not. A valid probability can still send software down the wrong branch.
The release contains three checkpoints rather than one universal model. The English and typed-decisions variants use 421-million-parameter ModernBERT encoders; the multilingual checkpoint uses a 322-million-parameter mmBERT encoder. The Hugging Face model card carries an Apache 2.0 license, and the Python client is available through PyPI. The basic install is short:
pip install laya
The operational footprint is less tiny than that command suggests. Loading all three checkpoints takes about 4.6GB in 32-bit precision, according to the repository's memory notes. The default router holds one model and evicts the least recently used one. Alternating languages with a cold model can therefore add four to six seconds on the CPU setup tested by the project. Preloading avoids that swap but keeps more weights resident. Self-hosting removes a metered API bill and moves the compute bill to the operator.
What the 33ms claim measures
Laya's headline latency came from a Tesla T4 run. The multilingual checkpoint answered one question in 32.8ms, while the English checkpoint took 39.5ms. Ten questions took 72.3ms on the multilingual model, or 7.2ms per question when batched. Those figures are in the project's benchmark report, along with links to the notebook and saved result files on the research branch.
The benchmark report compares those runs with published Jev measurements of 236 to 276ms at the median. It also supplies the warning that belongs beside the speed ratio: the Laya team had no Jev API access, and the two sets of prompts and sample sizes differ. Laya's claimed six-to-seven-times advantage is an indicative comparison across separate runs. It is not a controlled head-to-head test.
That caveat does not erase the local latency result. A bidirectional encoder can evaluate bounded questions without generating a token sequence, and several questions can share one pass. TypeSafe's own Jev announcement describes the same trade: possible outputs are defined in advance and sampled in parallel. Laya gives developers a way to test that design on their hardware, including the loading time and memory pressure that an API benchmark leaves outside the frame.
Confidence needs work after installation
The 51-language sweep shows why Laya has a router. On the MASSIVE intent test, where random accuracy is 0.050, the English checkpoint scored 0.000 on Khmer, 0.060 on Hebrew and 0.100 on Hindi. Its confidence stayed high in all three cases. The multilingual checkpoint improved Khmer to 0.180, Hebrew to 0.400 and Hindi to 0.430. Across the full sweep, it cleared three times random accuracy in 45 of 51 languages, compared with 23 for the English model, according to the per-language results.
Confidence alone cannot detect an alphabet the English tokenizer handles badly. Laya inspects Unicode scripts and some Latin-language cues before inference, then routes the request to a checkpoint. That is a sensible repair for the measured failure, but the routing documentation calls Latin detection a best effort. Teams serving mixed-language traffic still need examples from the languages and scripts their users send.
Calibration also requires a post-install step. Expected calibration error, where lower is better, was 0.466 for the English checkpoint as shipped. Fitting a temperature on held-out data cut the project's result to 0.081. The multilingual checkpoint moved from 0.314 to 0.106. The benchmark report says the multilingual weights ship without fitted temperatures. A developer who branches on a confidence threshold should fit and check that threshold against local data before automating an action.
The strongest typed-decisions score has another boundary. Laya reports 0.766 accuracy on 2,000 decisions for the checkpoint fine-tuned on that benchmark's training split. The two base checkpoints scored 0.361 and 0.342, both below the 0.461 majority-class baseline. The project's stated limits describe the base models as near chance on that task. This is a foundation to specialize, rather than a drop-in judge for an unfamiliar workflow.
Large option sets expose a separate ceiling. On Banking77, the specialized Laya checkpoint reached 0.492 accuracy, while a published Jev result listed in the report reached 0.870. Laya allocates a fixed token budget across the choices, leaving only a few tokens per label when the list grows to 77. The maintainers recommend keeping a choice question below about 20 options or splitting it into a coarse step followed by a narrower one. That constraint matters for teams hoping to replace a large routing table with a single model call.
Openness makes the weak results useful
The Laya repository includes the client code, tests, benchmark notebook and a research branch containing saved results. Its public Hugging Face entry exposes the 421-million-parameter English checkpoint as safetensors. These artifacts do not independently validate the launch numbers. They give other developers enough material to challenge them, which is the next test this release needs.
Some of the strongest application scores need that scrutiny. Laya reports 0.993 accuracy for email spam and phishing, but both datasets were in its training mix. Its held-out guardrail result topped out at 0.762, and held-out moderation was much weaker at 0.530. The application table labels those splits instead of blending them into one average. Anyone evaluating the model should preserve that separation and add a private test set from the intended workflow.
Watch next for independent reruns of the T4 latency and calibration tests, along with a release that ships fitted temperatures or improves option-order stability. The 794-point burst established that developers are curious about a learned branch that runs locally and returns typed probabilities. The worthwhile follow-up is less visible: take one real routing task, keep its labels out of training, and see how often Laya is confidently wrong before letting its answer move anything.