Qevi-2B: A Jev-style finetuned model for image classification

Products·launched

peekaboolean — 500M VLM answering typed image questions (choice/score/yes-no) with calibrated probabilities

peekaboolean is a small vision-language model that answers typed questions about a single image — multiple choice, rubric scoring, or yes/no — and returns calibrated probabilities instead of generated text. It solves the problem of slow, unparseable, poorly calibrated generative VLM outputs on local machines, for developers building apps that need fast image decisions. It runs about 400 ms per image on an M1 Pro without any text generation.

peekaboolean is a small vision-language model that answers typed questions about a single image — multiple choice, rubric scoring, or yes/no — and returns calibrated probabilities instead of generated text. It solves the problem of slow, unparseable, poorly calibrated generative VLM outputs on local machines, for developers building apps that need fast image decisions. It runs about 400 ms per image on an M1 Pro without any text generation.

A hobbyist LoRA fine-tune of Qwen3.5 4B that mimics Jev's behavior, trained on public datasets plus ~25M tokens of synthetic data generated by DeepSeek V4.1 Flash. The LLM is used as the assistant model itself, fine-tuned to improve typed-decision performance. It solves the problem of having an open, self-hostable Jev-like model and is released with dataset and API endpoint for local LLM users.

Products·launched

zico — open model that answers questions with probabilities instead of text

zico is an open model replicating Jev (TypeSafe AI's closed API): you describe a situation with questions and allowed answers, and it returns probabilities instead of text. It was trained in three days to run on a laptop and also supports image input. It targets developers who need small classification decisions inside apps.

zico is an open model replicating Jev (TypeSafe AI's closed API): you describe a situation with questions and allowed answers, and it returns probabilities instead of text. It was trained in three days to run on a laptop and also supports image input. It targets developers who need small classification decisions inside apps.

Releases·Tool Release

Dev reverse-engineers TypeSafe's Jev and releases open weights; Jev wins zero-shot on 7 benchmarks, finetuned clone wins on 395-label task

A developer who had paid for TypeSafe's Jev classification API probed it, reconstructed its design, and released an open-weights rebuild with a 37-page technical report. The closed Jev still wins zero-shot across all seven held-out benchmarks, but the open version wins on tasks with a few hundred finetuning labels and serves far more decisions per second on a single GPU. The release gives practitioners a local, deterministic alternative with full-precision probabilities.

A developer who had paid for TypeSafe's Jev classification API probed it, reconstructed its design, and released an open-weights rebuild with a 37-page technical report. The closed Jev still wins zero-shot across all seven held-out benchmarks, but the open version wins on tasks with a few hundred finetuning labels and serves far more decisions per second on a single GPU. The release gives practitioners a local, deterministic alternative with full-precision probabilities.

A Reddit user with early access to Typesafe.ai's Jev model ran it against a secret-detection benchmark suite built from gitleaks, GitGuardian, and similar sources. Results were described as promising and consistent, though the tester questioned whether Jev's low costs alone justify its use for this task.