We finally read the license. It wasn't the problem.
On July 13 we flagged one loose end on a 153 MB model we'd already rejected twice: we hadn't actually read its license, just noted it was "custom" and moved on. Today we read it. LFM2.5-230M's "lfm1.0" license turns out to be Apache-2.0 with one clause bolted on — free commercial use up to $10M in annual revenue, then you contact Liquid AI for a paid license — the same shape as Llama's own community license, which we already treat as acceptable. It was never the blocker we implied it might be.
So why is it still not in the catalog?
Because the actual answer was boring the whole time: age. LFM2.5-230M shipped June 25. Today is July 18. That's 23 days, nine past our 14-day freshness window, and freshness doesn't get a waiver for a license turning out fine. The engine gate cleared on July 13 (LFM2 and LFM2MoE are both in mainline llama.cpp, confirmed again today at tag b10066). The license gate clears today. The freshness gate is the only one left, and it's not negotiable — a model doesn't get to re-enter a 14-day window by us taking three weeks to read its paperwork.
$ curl -s https://api.github.com/repos/ggml-org/llama.cpp/releases/latest | grep tag_name
"tag_name": "b10066" # published 2026-07-17 23:13 UTC
$ grep -oE 'LLM_ARCH_[A-Z0-9_]+' src/llama-arch.cpp | sort -u | wc -l
137 # was 135 on July 13 (tag b9982)Two new architecture tags landed in the four days since our last engine check. Neither maps to anything in this week's model sweep, so it's a number worth logging and nothing more today.
What else we checked this week
Nothing else cleared the bar either, for more ordinary reasons. Gemma 4 E2B-it got a fresh GGUF quantization from ggml-org on July 16 — genuinely new upload, well inside our freshness window — but its smallest quant (Q4_0) is 2.84 GB, over our 2.5 GB phone-download cap. Thinking Machines Lab open-sourced Inkling, a 975B mixture-of-experts model, under Apache 2.0 on July 15 — real, well-documented, and roughly 400x too large for a phone regardless of quantization. Poolside's Ornith-1.0 family (9B/35B/397B) and a wave of 27B–284B community merges and roleplay fine-tunes from bartowski and unsloth rounded out the rest of this week's trending GGUF list. All outside our size budget; none evaluated further.
The catalog, unchanged
21 models, same as last check. No additions, no removals. Apple Intelligence remains the zero-download, zero-engine-version option on devices that support it.
One more thing worth a mention rather than a full post: this week's most interesting on-device story isn't a GGUF at all. A three-person research team just shrank a 27B model to 3.9 GB using a compression technique nothing in the llama.cpp ecosystem currently reads. We wrote up why that matters separately, below.
Apple's on-device AI moat, exposed by a 3-person Caltech lab it's now trying to buy →