signalHacker News Best2026-09-29
Jeff, Jev-compatible 0.8B decision models, trained at home, ~30 ms
Jeff is a family of small decision models fine-tuned from Qwen3.5 and Gemma 4 for zero-shot classification. It returns calibrated probabilities for each option in a single forward pass, with about 22 ms per decision on an RTX PRO 6000 and 28 ms on an Apple M4 Max. Trained entirely on local hardware without closed-model output, Jeff approaches Jev's benchmark performance on classification and grounding but lags on reasoning-heavy tasks.
- for who
- Developers needing fast, local zero-shot classification without large models.
- why now
- Jeff's new local decision models offer Jev-compatible speed for near-instant zero-shot classification, newly released.
- what changes
- They can embed a Jev-compatible classifier that runs on their own hardware, reducing latency and keeping data local.
- to do
- Run the provided uv commands to download and serve Jeff locally, then query it via curl.
key points
- Jeff models run ~22 ms on RTX PRO 6000 and 28 ms on Apple M4 Max.
- Fine-tuning on custom examples lifted accuracy from 31.7% to 95.8% in under half an hour.
- Jeff matches or beats Jev on classification tasks but lags on reasoning-heavy benchmarks.
#small models#zero-shot classification#local deployment#open source
score
score 9 out of 10. 0-10: how dense the facts are, multiplied by how much you can do with them after reading. 8+ means the topic's evidence bar is met: benchmarks and availability for a new model, amount and investors for a funding round, revenue figures for a solo-money story. Below 5 an item does not enter the digest. A press release scores 3 or less, a reprint loses 2, anything older than 14 days loses 1, a headline that misleads loses 3.
read the source