Jev Daily

Hassan's Tev1 0.8B classifies locally in about 50ms

A smaller sibling to tev1-4B, shown running on a laptop, with weights and benchmarks still to come.

Hassan trained Tev1 0.8B, which he describes as a tiny Jev-like classifier, and posted a video of it classifying tasks completely locally on his Mac through Ollama. He put end to end latency at roughly 50ms and noted the video is not sped up. Weights and benchmarks are promised soon but are not out yet.

Hassan
@nutlope
X
It's extremely fast: only ~50ms E2E latency. Video is not sped up!
Sep 24, 2026 · View on X
Hassan
@nutlope
X
Releasing weights & benchmarks very soon so you can try it yourself
Sep 24, 2026 · View on X

This follows the 4B model he put out earlier, tev1-4B-experimental, a Jev-like classifier finetuned on top of Qwen3.5 4B for $17, released with weights, the data recipe and a tutorial on training your own. That one is served on Together serverless at $0.042 per 1M input tokens and $0 per M output. The 0.8B version moves the same idea onto hardware you already own, which is the interesting part of the demo. There is no hosted endpoint in the loop, no network hop, and no per token bill.

What the 50ms figure is and is not

It is one number, from one person, on one laptop, running an unreleased model through Ollama on tasks he chose for a video. It is end to end, which is the number that matters for a decision layer sitting in front of a slower model, and it is the number a builder would want before wiring a classifier into a request path. It is not a benchmark, and it is not comparable to any published Jev latency figure, because nobody ran the two against the same workload.

The missing half is accuracy. Hassan has not posted a single evaluation score for Tev1 0.8B, which means the only measured property of this model right now is how fast it returns something. A sub billion parameter classifier can be fast and still be wrong often enough to make it useless as a gate, and the 4B release came with a data recipe and tutorial but the smaller one has nothing published yet to check quality against.

For people building on it

Do not plan a local fallback path around this until the benchmarks land. Speed is the easy half of a classifier and it is the half that has been demonstrated here. The useful thing to take from the post is the shape of the setup, a small Jev-like model on Ollama, no API call, latency low enough that it can sit in front of something expensive without the user noticing. Whether it is correct enough to do that job is an open question that Hassan himself has said he will answer with weights and benchmarks.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.