Releases·Tool Release
Jev at home, but it can see: typed yes/no, pick-one and rubric answers with per-label probabilities from Gemma 4 31B on a 4090, images included
A developer released typevet (MIT, Python 3.12), a library that gets Gemma 4 31B to answer yes/no, pick-one, and rubric questions with a probability per label — including images — locally on a 24 GiB RTX 4090 via llama.cpp, extending TypeSafe's text-only Jev approach. A receipt-verification test showed text-only answers confidently matched wrong totals that image input caught at ~0.99998 confidence.
5 engagement1 sourceTue, Sep 29, 2026, 09:38 PM
AI Overview
- typevet (MIT, Python 3.12) reads next-token distributions over allowed answer labels from Gemma 4 31B, adding image input Jev's docs don't support
- Receipt test on 6 CORD v2 receipts: swapped totals caught 0/6 text-only but 6/6 with photo, at 0.99998+ probability; masked totals abstained every time
- Images go to llama.cpp via base64 prompt.multimodal_data; token-count growth gates verify the image arrived
- On vLLM 0.30.0 with BF16 Gemma 4 31B on one H100: 18/18 receipt test, 0.24 s median per record, 39.6 records/s at 64 in flight
- Prior art: TypeLLM added image input 9/24; allanrbo posted a Gemma 4 12B webcam script 9/25; VQAScore reads Yes-probabilities since 2024
Related
Social·Community
Jev by TypeSafe: Building AI Systems That Make Decisions, Not Just Generate Text
Medium post examines Jev by TypeSafe as AI that makes decisions, not just generates text
Medium post examines Jev by TypeSafe as AI that makes decisions, not just generates text
0 engagement2 sourcesrssMon, Oct 12, 2026, 03:06 AM
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.
1 engagement1 sourceredditSun, Sep 27, 2026, 08:46 AM
Products·launched
hearim: open-source gateway turning ordinary local LLMs into a Jev-compatible decision API
hearim is an open-source gateway that lets existing local or hosted LLMs act as a Jev-compatible decision API. Given data and a set of choices, it returns a decision with probabilities without task-specific fine-tuning or custom weights. It targets developers running models via llama.cpp, vLLM, SGLang, Ollama, or OpenAI-compatible servers.
hearim is an open-source gateway that lets existing local or hosted LLMs act as a Jev-compatible decision API. Given data and a set of choices, it returns a decision with probabilities without task-specific fine-tuning or custom weights. It targets developers running models via llama.cpp, vLLM, SGLang, Ollama, or OpenAI-compatible servers.
1 engagement1 sourceredditMon, Sep 21, 2026, 07:59 AM
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.
0 engagement1 sourceredditSun, Sep 20, 2026, 04:31 PM
Products·building
semif-go: Go server giving local llama.cpp models a typed decision API
A Go server that lets local llama.cpp models answer small decisions — yes/no probabilities, typed choices, scores — using next-token probabilities instead of generating paragraphs or JSON to parse. It targets developers running local model harnesses who want lightweight structured answers. Tested with Gemma 4 12B for text and images.
A Go server that lets local llama.cpp models answer small decisions — yes/no probabilities, typed choices, scores — using next-token probabilities instead of generating paragraphs or JSON to parse. It targets developers running local model harnesses who want lightweight structured answers. Tested with Gemma 4 12B for text and images.
20 engagement1 sourceredditWed, Sep 23, 2026, 09:11 PM