Entity trail

Jared Palmer

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-20 / skill-finderkev trains a typed-decision model on a MacBook in 1h45m and lands within 6-7 accuracy points of the commercial referencejaredpalmer/kev (created 2026-09-17, 733 stars, pushed 09-20) is a LoRA adapter plus a custom readout head on Qwen that answers many typed questions about one document in a single forward pass, across yes/no, multiple-choice and ordered-scale question types. Training kev-0.5b takes about 1 hour 45 minutes on an Apple M5, larger variants 40-70 minutes on one H100, using ~1,000 records each from 10-13 public sources (Banking77, AG News, MNLI, BoolQ, SST-5, Yelp) plus synthetic policy pairs. On held-out out-of-domain sources kev-4b scores 0.790 accuracy / 0.328 Brier versus the Jev reference at 0.857 / 0.211, serves in about a second on a 32GB Mac in bf16, and speaks the official typesafe-sdk wire format so you can point an existing client at localhost.
2026-09-19 / thought-leaders-researcherJared Palmer's Kev-0.5B puts a Jev-shaped decision model on a MacBook ProVercel's Palmer built a LoRA adapter plus a small readout head on Qwen2.5-0.5B that runs locally on a MacBook Pro, one of the six clones catalogued by AINews. The interesting part is who built it: Vercel is simultaneously shipping Jev as a hosted gateway model and demonstrating a laptop-sized substitute for it. That is a hedge against the exact lock-in risk a closed-weights decision model creates for anyone putting a classifier in an agent's hot path.
2026-09-19 / thought-leaders-researcherSix independent clones of TypeSafe's closed Jev model shipped within two days of launchLatent Space's AINews catalogued six reimplementations of Jev built in 48 hours: Laya (421M params, ModernBERT-large encoder, Nandha Kishor M), DiffusionGemmaJev (a vLLM pull request), Bespoke Nimble (@madiator's LoRA on Qwen3.5-9B), SemIf/OpenJev (Alex Wortega, Qwen3.5 backbone with a three-class NLI head), Jevlike (Vinny La Rouge, a 40KB byte-embedding option-attention model) and Kev-0.5B (Jared Palmer, a LoRA adapter on Qwen2.5-0.5B that runs on a MacBook Pro). The trigger was that Jev is closed weights, so the architecture had to be guessed from behavior. Harrison Chase's pick for the best application so far is browser-use agents, and @ankrgyl reports roughly 400x lower scoring cost than his prior setup.

Graph relationships

WORKS AT
Vercel
Jared Palmer is at Vercel
Source finding
BUILT BY
Kev-0.5B
Jared Palmer built Kev-0.5B, a LoRA adapter on Qwen2.5-0.5B for MacBook Pro
Source finding
BUILT BY
Kev-0.5B -> Jared Palmer
Jared Palmer built Kev-0.5B, a LoRA adapter on Qwen2.5-0.5B for MacBook Pro
Source finding

Source trail

GitHub GitHub GitHub GitHub GitHub Hugging Face Latent Space (swyx)

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…