choosekit runs Jev-style local decisions via llama.cpp: 96.53% accuracy, 239 ms vs Jev's 368 ms

Social·Community

Reddit user replicates JEV locally with a 1B model, matching ~90% of its performance on a laptop CPU

A Reddit user reproduced the JEV setup (text plus question in, P(yes) out) with a locally trained 1B model running offline on a laptop CPU with no API calls. The local approach runs roughly 20x faster than JEV once network latency is included and reaches about 90% of JEV's accuracy on 999 hand-written yes/no questions, suggesting the underlying capability is not exclusive to the hosted model.

A Reddit user reproduced the JEV setup (text plus question in, P(yes) out) with a locally trained 1B model running offline on a laptop CPU with no API calls. The local approach runs roughly 20x faster than JEV once network latency is included and reaches about 90% of JEV's accuracy on 999 hand-written yes/no questions, suggesting the underlying capability is not exclusive to the hosted model.

Products·concept

JevRev — a semantic decision layer that scores and filters LLM plans before more tokens are spent

JevRev pairs an LLM with 'Jev', a semantic layer that shortlists options, scores them in a loop, and monitors progress on ongoing work. Instead of being another coding agent, it acts as a decision layer beside the LLM, filtering plans and checking attention so cheap routing decisions aren't left to the model alone. It targets developers using LLMs for tasks that generate many candidate approaches, preventing wasted time and tokens on low-value directions.

JevRev pairs an LLM with 'Jev', a semantic layer that shortlists options, scores them in a loop, and monitors progress on ongoing work. Instead of being another coding agent, it acts as a decision layer beside the LLM, filtering plans and checking attention so cheap routing decisions aren't left to the model alone. It targets developers using LLMs for tasks that generate many candidate approaches, preventing wasted time and tokens on low-value directions.

Products·launched

jevos: fast yes/no classifier model that runs on laptop CPU

jevos is a small model for binary yes/no questions: you feed it text and a question, it does a single forward pass and returns a probability of yes. It replaces calling a full LLM for trivial decisions, runs on a regular laptop CPU at ~54ms per query, and exposes a JEV-compatible API so existing JEV users can switch to it.

jevos is a small model for binary yes/no questions: you feed it text and a question, it does a single forward pass and returns a probability of yes. It replaces calling a full LLM for trivial decisions, runs on a regular laptop CPU at ~54ms per query, and exposes a JEV-compatible API so existing JEV users can switch to it.

Developments·Benchmark

Opper benchmarks Kev 4B vs. Jev: matching accuracy, weaker calibration, up to 12x higher short-request cost

Opper ran Jared Palmer's Apache-2.0 Kev 4B, a fine-tune of Qwen3.5-4B, side by side with Jev on the same endpoint using 362 items published after both models shipped. Accuracy landed within 2 points on every task, though Jev showed better calibration and stronger paraphrase detection. Cost differences favored Kev on short requests because Jev adds roughly 257 fixed extra input tokens per request.

Opper ran Jared Palmer's Apache-2.0 Kev 4B, a fine-tune of Qwen3.5-4B, side by side with Jev on the same endpoint using 362 items published after both models shipped. Accuracy landed within 2 points on every task, though Jev showed better calibration and stronger paraphrase detection. Cost differences favored Kev on short requests because Jev adds roughly 257 fixed extra input tokens per request.

Social·Community

knoopx runs Jev-style decisions API natively on ninfer with Swift-Qwen3.8-27B, scoring 84.4% on JevBench v1.2 public set

A developer built a proof-of-concept jev-like API natively on ninfer, mix-and-matching open implementations on top of Qwen3.8-27B, and benchmarked it against JevBench v1.2. The self-hosted run hit 84.4% public-set accuracy with strong calibration, near top published systems' accuracy though with no published composite score. The author calls it a dirty POC with remaining bugs and allocation issues.

A developer built a proof-of-concept jev-like API natively on ninfer, mix-and-matching open implementations on top of Qwen3.8-27B, and benchmarked it against JevBench v1.2. The self-hosted run hit 84.4% public-set accuracy with strong calibration, near top published systems' accuracy though with no published composite score. The author calls it a dirty POC with remaining bugs and allocation issues.