Hassan open sourced Tev1 0.8B, a Jev-like classifier for laptops
The weights are out, the promised benchmarks are not, and Hassan says it is worse than both Tev1 4B and Jev.
Hassan has released the weights for Tev1 0.8B, a tiny Jev-like classifier that runs entirely on a Mac. It follows his earlier Tev1 4B, the Jev-like classifier he trained for $17 and open sourced.
Definitely not as good as Tev1 4B or Jev, but it runs entirely locally on a mac.
The demo came first. In his earlier post he showed the model running locally through Ollama and classifying tasks at what he measured as roughly 50ms end to end, with a note that the video was not sped up. That is the only latency number attached to the model so far, and it is a screen recording on one machine, not a benchmark run.
What Hassan says it is not
The release post is unusually direct about the ceiling. Hassan writes that it is "definitely not as good as Tev1 4B or Jev" and that it "can work decently well for very simple classification." That is the vendor of the model talking down his own model, which is worth more than most eval tables, but it is still a qualitative claim. There is no accuracy figure, no task list and no comparison run against either Tev1 4B or Jev.
When he first posted the model he said weights and benchmarks were both coming very soon. The weights arrived. The benchmarks have not. So the two things a reader would want to put next to each other, the 50ms and the hit rate, currently exist only one at a time.
For people building on it
The pitch here is not quality, it is where the model sits. At 0.8B parameters it runs on the laptop, which removes the network hop and the per call price from the classification step entirely, and it removes the dependency on an API that has paused new signups. That is a real structural difference for anyone doing high volume, low stakes routing.
But "very simple classification" is doing load bearing work in that sentence, and the failure modes of a model this size on anything ambiguous are not documented yet. The pattern that has worked elsewhere is cascading, sending the cases a small model is unsure about to something bigger, which is roughly what Hassan himself did when he routed Jev's low confidence calls to Kimi K3. Nothing in either Tev1 0.8B post says the small model exposes usable confidence scores, so if you want that shape you should check before you plan around it.
Until the benchmarks land, treat the 50ms as a demo figure and the accuracy as unknown. Download it, run your own labeled set through it, and assume it loses to the 4B on anything with nuance, because the person who trained both says it does.
It's extremely fast: only ~50ms E2E latency. Video is not sped up!