Jeff

Jev at home, with the receipts to prove it was done at home. Jeff is three open-weight fine-tunes (Qwen3.5-0.8B, Qwen3.5-2B, Gemma4-E2B) that do zero-shot classification in the same request format as Jev: hand them a situation in plain words plus any list of options, hand it a situation and your options, it answers back with a confidence score on each option in one look, about 22 ms on an RTX PRO 6000 and 28 ms on an Apple M4 Max. Apache 2.0, independent project, not affiliated with TypeSafe; the training code starts from their open-source AutoJev recipe.

The benchmark picture, both halves. On the five-benchmark panel the 2B scores 83.1 against Jev’s published 83.0 - a tie - and Financial PhraseBank (96.3 vs 77.0) and RAGTruth (88.9 vs 77.3, tied with the 27B AutoJev) go to Jeff outright. The reasoning-heavy rows do not: BBH 68.0 vs 94.3, WinoGrande 79.0 vs 90.7, and JevBench’s hard tier 53.3 vs 73.3. Zero-shot classification and grounding at 2B match a flagship; reasoning at 2B does not. An early production comparison from an HN commenter landed the other direction entirely: 70% vs Jev’s 94% on their own routing task. Both are true, and the difference is workload: benchmark-typical classification holds, out-of-distribution demands the fine-tune - which the author demonstrates by moving held-out accuracy from 31.7% to 95.8% in under half an hour on one GPU.

Everything in the making was local. Trained on one RTX PRO 6000 (0.8B in ~2 hours, 2B in ~3.5), synthetic training data written by open Qwen3.8-Flash-Next on two DGX Sparks, tested on a MacBook - no cloud GPUs, and no closed-model output in the training data. The training bill is two hours of one workstation GPU. That is the real headline: the recipe is reproducible with the hardware that is already in the house.

Limits, read before deploying. The released models handle at most 26 options per question - options are letter-coded A-Z, and anything past position 26 is effectively never chosen (the server refuses the request; 255-option retraining is in progress). At 0.8B-2B parameters, none of these reason multi-step, for any model in class. And benchmark scores do not predict game play: the model that wins the benchmark suite lost the games to a smaller sibling.

Where this sits in the family map. Jev (TypeSafe) made the decision-model category credible at flagship scale; AutoJev reproduced it open at 27B; GLiNER2.5-Decide covers the CPU-class fixed-schema niche; Jeff is the smallest reproducible end of the spread - a decision model that trains in an evening, ships under 8 GB of RAM, and returns a choice for about nothing. For the routing-and-gating calls that currently burn frontier tokens, the answer is now installable rather than subscribed.