At the end of the last article, I said I wanted to try a model small enough to run inside Firefox itself, for people who don’t run Ollama. That model exists, and it runs. The ordinary release of Firefox on my laptop, with no settings changed, answered Xift’s four questions about a post on its own. Nothing went to a server, and no key or account was needed.
It took one to two seconds a post. And outside two of the four questions, the answers weren’t good enough to decide what my feed collapses.
The model is kev-0.6b, the default in open-jev, a small library that runs “open reproductions of the shape” of Jev in the browser on top of Transformers.js. It has 0.6 billion parameters and runs on the graphics chip through WebGPU.
I didn’t touch Xift itself. I built a throwaway add-on that loads the model and times it on posts I wrote.
Speed was the surprise.
Most of that second isn’t the model. One question took 959 ms and four took 1,070, so nearly all of it is a fixed cost on every call. Mozilla has tracked the gap since June 2025, in a bug titled “qwen3-webgpu is ~10x slower than Chrome”.
Real posts take longer than mine. On the 442 live posts I use below, the median was 1.7 s a post.
It also runs only where Firefox has WebGPU. kev-0.6b was the only one of open-jev’s two models that worked, and the fallback that runs on the processor couldn’t load the compressed version of either.
| Platform | WebGPU in Firefox | kev-0.6b in Firefox |
|---|---|---|
| Apple silicon Mac | Yes | Runs (tested) |
| Windows | Yes | Should run (not tested) |
| Linux | No | Doesn’t run |
| Intel Mac | No | Doesn’t run |
Then the answers. I gave it the 442 posts from the last article, everything two ten-minute scrolls of X searches showed me, and compared its scores with Jev’s. Scores don’t carry across models, so I checked order: sorted by the model’s score, do the posts Jev flagged come first? I asked the same of the two Ollama models, on the same posts.
| Question | Posts Jev flagged | kev-0.6b, Firefox | Tev1 0.8B, Ollama | Tev1 4B, Ollama |
|---|---|---|---|---|
| Rage bait | 7 | 0.81 | 0.81 | 0.99 |
| Self-promotion | 46 | 0.72 | 0.76 | 0.98 |
| Low substance | 155 | 0.48 | 0.57 | 0.89 |
| Engagement bait | 13 | 0.46 | 0.68 | 0.70 |
On rage bait and self-promotion, kev-0.6b ranked about as well as Tev1 0.8B. On engagement bait and low substance, it did no better than chance. Tev1 4B came closest to Jev on all four, but it’s also the slowest.
That matches what the model’s author says. Its model card says “Out of domain it is a 0.6B model”, right about 62% of the time on questions unlike its training data, where Jev gets 86% on the same tests.
There’s a wait before the first answer, too. The model is 346 MB, more than the 200 MB that Mozilla’s add-on site accepts, so it downloads on first use. That took me almost five minutes.
So a decision model does run inside Firefox, with nothing to install, no key, and nothing I read leaving the browser. But it takes one to two seconds a post, does no better than chance on two of the four questions, and doesn’t run at all on Linux or Intel Macs. That isn’t ready to be a setting in Xift. Local stays Ollama, and Jev stays the default.
Three things could change that. Mozilla is working on the delay: the change that bug waits on, “Don’t poll WebGPU from a timer”, has an owner. Kev 1.0 has a 0.8B model, which kev.js runs in Chrome; I haven’t tried it in Firefox. And Chrome is trialling a built-in Decisions API, with the model inside the browser itself, behind a flag in Chrome 157. Mozilla hasn’t taken a position on it.
Until one of those lands, the quickest way to run a decision model for Firefox is on the laptop beside it, through Ollama.