hearim: open-source gateway turns ordinary local LLMs into a Jev-compatible decision API
ziozzang0 · reddit · 2026-09-21
A developer released hearim, an open-source gateway arguing you don't need a specialized model for Jev-style decisions — ordinary local or hosted LLMs suffice.
How it works:
- Turns a decision into a constrained set of alternatives and reads the pretrained model's probability distribution directly, instead of parsing generated JSON and hoping the schema holds
- Returns decisions with probabilities via the Jev Choice, Score, and Noul API format
- Works with models served through llama.cpp, vLLM, SGLang, Ollama, or any OpenAI-compatible server
Core claim: pretrained models already learned a lot before instruction tuning/RLHF; if you only need a choice among clear alternatives, much of that capability may already exist in the probability distribution — "the model may already know. We just need to stop asking it to write an essay."
The author notes the underlying technique predates Jev's announcement (used privately for over a year); hearim is an independent reimplementation around the Jev API contract and makes no claims about reproducing Jev's architecture, training, calibration, or speed. Repo includes a paper, supported backends, limitations, and benchmarks. He's especially curious about results with smaller base models, not just instruction-tuned ones.
More from coding & agent
- Portracker: open-source self-hosted tool auto-discovers running services and network ports — tom_doerr · 2026-09-21
- Termcp: AI-native terminal platform lets agents drive real interactive CLIs without per-tool MCPs — Psychological-Bid722 · 2026-09-21
- 'Software engineering is dead'? Dev lists databases, K8s, LLM ops that English won't replace — ashishllm · 2026-09-21
- How to turn Claude into your first employee before you can afford one — evielync · 2026-09-21
- Jev guide: the System-1 model that only judges and never writes — xiaohu · 2026-09-21
- Opinion: If an agent can't buy your product, you're building last year's SaaS — diegoposts · 2026-09-21