Releases·Tool Release
Brio mode in Colibrì 1.12.0: score allowed answers instead of generating a reply
Colibrì founder Vincenzo announced Brio mode in version 1.12.0, which scores a provided set of candidate answers instead of generating a reply. Users supply context, a question, and possible answers; the tool returns a probability for each option plus the entropy of the distribution, using locally supported models without new training.
1 engagement1 sourceMon, Sep 21, 2026, 11:52 AM
AI Overview
- Colibrì founder Vincenzo released Brio mode in 1.12.0, inspired by TypeSafe, returning per-option probabilities and entropy without generating output tokens
- Brio works via dashboard, terminal, or POST /v1/brio, reusing shared context through snapshots for multiple questions per document
- Founder warns Brio's normalized scores aren't calibrated confidence—low entropy favors an option but doesn't guarantee correctness
Related
Products·launched
OpenJev: Qwen3.5 4B logit-probability scoring replicating Jev-style choice models
OpenJev is an open implementation showing that Qwen3.5 4B with per-option logit probabilities can match or beat TypeSafe AI's Jev model, which outputs probabilities over choices. It offers a browser demo for trying it directly. It targets local LLM users wanting cheap multiple-choice probability scoring without a dedicated reranker.
OpenJev is an open implementation showing that Qwen3.5 4B with per-option logit probabilities can match or beat TypeSafe AI's Jev model, which outputs probabilities over choices. It offers a browser demo for trying it directly. It targets local LLM users wanting cheap multiple-choice probability scoring without a dedicated reranker.
68 engagement1 sourceredditWed, Sep 16, 2026, 03:10 PM
Releases·Tool Release
LLM2Jev turns local LLMs into Jev-style structured decision engines with prefill-only binary inference
A new open-source project, LLM2Jev, adapts local Hugging Face-compatible causal language models into Jev-style structured decision engines, emitting typed Choice, Score, and Noul answers with probabilities. It uses prefill-only binary inference — scoring candidates with yes/no logit judgments without generating answer tokens — and leverages SGLang's Radix Cache for shared-prefix efficiency. The project is independent and unaffiliated with Jev or TypeSafe.
A new open-source project, LLM2Jev, adapts local Hugging Face-compatible causal language models into Jev-style structured decision engines, emitting typed Choice, Score, and Noul answers with probabilities. It uses prefill-only binary inference — scoring candidates with yes/no logit judgments without generating answer tokens — and leverages SGLang's Radix Cache for shared-prefix efficiency. The project is independent and unaffiliated with Jev or TypeSafe.
109 engagement1 sourcegithubSat, Sep 19, 2026, 10:59 AM
Products·building
AES: a local Jev-style decision head on Qwen 2.5 1.5B for non-autoregressive decisions
AES is a proof-of-concept that attaches a Jev-inspired decision head to Qwen 2.5 1.5B Instruct, testing whether decisions can be made without autoregressive token generation. It runs a single prefill pass, extracts hidden states via a custom probe, and compares them to candidate token embeddings. It targets local open-weight users exploring non-autoregressive decision engines.
AES is a proof-of-concept that attaches a Jev-inspired decision head to Qwen 2.5 1.5B Instruct, testing whether decisions can be made without autoregressive token generation. It runs a single prefill pass, extracts hidden states via a custom probe, and compares them to candidate token embeddings. It targets local open-weight users exploring non-autoregressive decision engines.
6 engagement1 sourceredditSun, Sep 20, 2026, 11:06 AM