Capabilities and workflow
| Criterion | Jev (TypeSafe)Source | Laya (Convai Innovations)Source | StayCharted AI Model TrainerSource |
|---|---|---|---|
| What it is | A hosted "System One Model": text in, typed decisions out | An open-source typed-decision model family with English, multilingual and fine-tuned checkpoints. | A managed, no-code trainer for private classification models, with review |
| Maker | TypeSafe | Convai Innovations | StayCharted |
| How you get it | TypeSafe API, OpenRouter and developer integrations such as Mastra. | Download it: Python package, a self-hosted server in Jev's format, MCP and LangChain add-ons | In the browser, plus a REST API; AWS and Microsoft Marketplace |
| Starts from | A question and its options; no examples | A question and its options; best after fine-tuning | Labeled examples; on Essentials and Business, suggested groups and labels from unlabeled rows for your team to review. |
| Trains on your examples | No: shaped by instructions and criteria | Yes, with a notebook you run (two T4 GPUs) | Yes, built in |
| Answer types | A choice, a score, a yes/no; several questions per request | A choice, a score, a yes/no | A category, several fields at once, or a list, with a confidence |
| Inputs | Text; default 64K tokens per request and 32K per question. | Text (English, plus a multilingual checkpoint) | Text columns and permission exports; pictures with Vision on Essentials or Business. |
| Languages | Best in English | English; multilingual checkpoint (100+ claimed; 45 of 51 tested usable) | Trains on your examples in their own language; test on your data |
| Speed (maker's figures) | About 70–500 ms end to end | Project-reported: 32.8 ms multilingual / 39.5 ms English for one question on a T4. | Files in batches or REST API calls; measure latency with your own model and workload. |
| Confidence | TypeSafe describes calibrated probabilities; validate on your workload. | A confidence; the card recommends refitting it on your data | A confidence with every answer, and a review cutoff you choose |
| Accuracy you can check | You test it on your own sample | You test it; the card publishes its own benchmarks | Held-out accuracy, per category, for every version |
| Review and retraining | Up to your application | Up to your application | Review Queue on Essentials and Business; corrections feed the next training version. |
| Who builds it | Developers | Developers | Business teams; developers through the API |
| Cost | $0.042 per million input tokens; output free (checked 20 October 2026) | Free (Apache 2.0); your own hardware | Free plan; Essentials $49, Business $99 per workspace a month |
Fit by requirement
Hosted decisions in code
Evaluate Jev when your team wants an API without operating a model server.
Open source on your servers
Evaluate Laya when your developers can manage fine-tuning, calibration, security and hosting.
No labeled examples yet
Use Jev or Laya for draft labels and verify them, or use StayCharted’s grouping workflow on Essentials and Business.
Your organization’s categories
Evaluate a model trained on reviewed decisions for assignment groups, GL codes, product categories or access suggestions.
Measured accuracy and review
All three can be evaluated. StayCharted includes held-out reports and, on Essentials and Business, a Review Queue. With Jev or Laya, build or integrate these workflows.
Photos, or a photo with a note
StayCharted supports pictures through Vision on Essentials and Business. Jev and Laya’s documented inputs are text.
Buying through cloud marketplaces
StayCharted is available through AWS Marketplace and Microsoft Marketplace.
What the published benchmarks say
None of these were measured in the same test. The Jev number is a third-party result cited by Laya; Laya and StayCharted report their own tests.
Laya's model card, on Banking77 (77 intents):
- 0.425 for Laya, at default settings;
- 0.870 for a third-party published Jev figure using 72 labels.
Laya's card says Jev's numbers were "never measured here", with different samples and prompts. TypeSafe's own launch post and docs, as checked, don't publish a Banking77 result.
Laya's card, on other tasks: it reports beating Jev's published figures on some, such as news topics (AG News) and emotions. Zero-shot, its base checkpoints are close to chance on its typed-decision test; fine-tuning takes that from 0.362 to 0.766.
StayCharted, on Banking77: trained on the examples and tested on held-out rows, 92.6% (AI Classifier Model) and 93.7% (Dedicated AI Model). That's a trained setup, unlike the zero-shot figures above. The banking benchmark
Sources: Laya model card and StayCharted’s BANKING77 method. No controlled three-way comparison has been run by StayCharted.
Before choosing
- Set aside a few hundred items your team has already decided. Keep these test items out of all training data.
- Ask Jev the question with your categories as the options.
- Run Laya as it comes, and again after fine-tuning on separate training examples if you have the hardware.
- Train a StayCharted model on those same training examples.
- Compare all three on the untouched test items, category by category. Count data preparation, hosting, calibration and review work alongside accuracy, cost and latency.
Use the same category definitions and independently reviewed test labels for every approach.
Sources
TypeSafe model documentation · checked October 20, 2026TypeSafe launch · checked October 20, 2026Laya model card · checked October 20, 2026Laya README · checked October 20, 2026OpenRouter Jev listing · checked October 20, 2026Product names belong to their respective owners. StayCharted is not affiliated with the vendors compared here.
StayCharted plans and workflow
Free covers text and includes up to one AI Classifier Model. Essentials is $49/month and Business is $99/month per workspace. Pictures require Vision: +$49/month on Essentials or +$99/month on Business. Dedicated AI Models require Business. Prices are USD.
StayCharted trains and runs its own models, or uses your company’s own AI for text classification. With the Your AI Models add-on ($49/month on Essentials, $99 on Business), it uses your AI on Amazon Bedrock, Microsoft Foundry, Google Cloud, OpenAI or Anthropic, gives it your categories and most similar past examples, and tests it on rows it hasn’t seen. On Business, Dedicated AI Models can also be trained in your own Amazon Bedrock; the trained model stays in your AWS account. Your own AI options
- Every training run reports accuracy on examples not used to train the model, overall and by category. Answers include confidence.
- Data-quality checks and personal-information masking before text training are available on every plan.
- On Essentials and Business, answers below your review cutoff, from files, Google Sheets, the REST API (including Zapier, Make, Power Automate and ServiceNow) and AI assistants, wait in one Review Queue; reviewed answers go into the next version’s training. Grouping unlabeled examples is also included; pictures require Vision.
- Picture checks flag near-duplicates and potentially mislabeled examples. Location metadata is stripped from the copies used by the model; original uploads are retained unchanged until deletion.
StayCharted is a hosted service: there’s no on-premises version, and it can’t import models trained elsewhere. A Builder can download a trained Dedicated AI Model (its adapter, which needs the exact base model and your own inference software), an AI Classifier Model or an AI Image Classifier Model from the model’s Manage tab. A Classifier Model runs only inside StayCharted, and a model trained in your own Amazon Bedrock is already in your AWS account.
Compare plans and allowances · Model options and downloads · Data handling
Common questions
What's the difference between Jev and Laya?
Jev is TypeSafe's hosted API. Laya is an independent, open-source model you run yourself, whose server accepts Jev's request format. Laya can be fine-tuned on your data; TypeSafe doesn't fine-tune Jev on customer data.
Is Laya an open-source Jev alternative?
Yes. It's Apache 2.0, and works with code written for Jev's API. Its model card says it does best after fine-tuning and calibrating it on your own data.
Is Jev or Laya more accurate?
The figures cited here use different datasets, samples, prompts or label counts and do not establish a head-to-head winner. Test the same model versions on the same held-out examples.
When should I train my own classifier instead?
When the categories are your organization's own, you have past decisions to learn from, and you need measured accuracy and a review step. StayCharted does that without code.
Do Jev or Laya classify pictures?
Both document text inputs. StayCharted supports pictures and pictures with notes through Vision on Essentials and Business.
Can I use Jev or Laya with StayCharted?
Not built in. You can label rows with either, have your team check them, and upload them to train a StayCharted model.