Jeff: open decision models you fine-tune at home, matching Jev

Apache-2.0 weights hit 83% on a single GPU in hours; plus a 1.2B open OCR model that tops OmniDocBench, and tiny LLMs that run fully in your browser.

Nowline SEP 29 8:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • 83% accuracy from a classifier you trained yourself

    Jeff's 2B Qwen fine-tune scores 83.1% across five decision benchmarks — dead level with the commercial Jev API's 83.0% — while the 0.8B reaches 79.1%. Weights ship Apache-2.0 and the code is MIT, and training used only synthetic data from open models, so nothing stops you shipping it in production.

  • Two hours on one GPU, then ~22ms per call

    A full fine-tune runs about 2 hours for the 0.8B and 3.5 for the 2B on a single RTX PRO 6000. At inference it returns a calibrated probability per option in ~22ms on that card and ~28ms on an Apple M4 Max — cheap enough to drop a classifier straight into support routing, moderation, intent detection, or game AI.

  • The open decision-model tier is filling out fast

    Jeff lands the same week as fastino's GLiNER2.5-Decide (0.5B, CPU) and AlexWortega's openjev — a cluster of open, calibrated classifiers aimed squarely at paid decision APIs. If you've been renting Jev-style routing per call, there's now a self-hosted swap you own.

  • TeleOCR: a 1.2B open doc parser that beats the pack

    XingChen-AGI's TeleOCR (Apache-2.0, 1.2B) posts 96.87 on OmniDocBench v1.6 — state-of-the-art, ahead of MinerU 2.5 Pro and PaddleOCR-VL — and reads distorted, camera-shot pages without a separate rectification step. It runs on Transformers, vLLM, SGLang or Docker; wire it into a document pipeline today.

  • Run tiny LLMs in the browser, zero API bills

    MicroLLM Lab loads 25M–360M models (SmolLM2, PetitGPT and friends) via WebGPU, Q4-quantized to 50–84MB, with sub-10ms time-to-first-token and nothing leaving the device. It's a quick weekend build for on-device query classification or spam filtering before you pay for a cloud call — open source, cached in IndexedDB.