The trending local models nobody has measured

There is a gap between "popular on Hugging Face" and "someone has actually run it at home." We pulled the trending text-generation list and cross-checked it against every run on our board — 468 rows, 165 models, 70 GPUs as of 2026-10-07 (the model count fell from 239 in September because duplicate entities were merged, not because rows were removed).

What we found

Of the top 40 trending text-generation models, 27 have at least one run on the board. The other 13 mostly fall into two buckets.

Bucket 1 — too new for anyone to have run. The clearest case is the IFM/K2-Horizon family (several sizes, including a ~36B A4B MoE). It appears repeatedly in the trending list and has zero community runs in the upstream data we import from. Not "we missed it" — nobody has published a measured run yet.

Bucket 2 — a smaller sibling got measured instead. openbmb/MiniCPM5-2B is trending; the run we have is its 1B sibling. Same story for a few others where the 1–3B variant has numbers and the larger one doesn't.

A handful of entries are simply out of scope for a hardware board: a TTS model, and gpt2 — a baseline, not something you buy a GPU for.

Why this happens

New MoE models at 30B+ are exactly the ones that need a specific kind of rig to run well: enough VRAM to hold the active experts, a recent engine build, and a quantisation that keeps quality. That combination lags the release by days to weeks. Meanwhile the people with that hardware are busy running whatever they already have.

So the gap is not a data-collection failure. It is a hardware-and-tooling lag, and it is the most useful thing a benchmark site can point at.

What we're doing about it

• We keep importing the runs that do exist, attributed per row, every week — the board has grown from 316 rows in early September to 468, mostly by closing coverage gaps like these.
• Every model page shows what we have and, by omission, what we don't. If a card or model has no runs, you can see it.
• The submit flow exists precisely for the models the upstream data misses: one prompt to your local model, three steps, no downloads. If you have the hardware for a K2-Horizon-class model, a single run fills a hole that the whole community is currently guessing about.

The takeaway

Trending is not the same as measured. Before you plan a build around a model that is everywhere on the model hub, check whether anyone has posted a real number for it. If they haven't, the honest answer is "unknown" — and that is worth more than a made-up one.

Sources

• Trending check: Hugging Face text-generation models by trending score, read 2026-09-10; per-model run counts from the localmaxxing public API the same day.
• Board state: /benchmarks · methodology
• Coverage detail (internal): docs/HF_COVERAGE_2026-09-10.md