SLM Daily #16: llama.cpp's first real GitHub Release landed with a 280B-parameter model — and nothing phone-sized
llama.cpp cut v0.2.0 on August 21 — the first v0.x.x tag to appear on the project's Releases page marked "Latest" rather than as a bare pre-release. Per the project's own release docs, semver tags have existed for a while (v0.1.2 shipped August 18), but historically "releases are not published to github releases, only nightly/development builds are available there" — the tag itself was the artifact, with no Release entry behind it. v0.2.0 is the first one to actually get that page. We're not calling this "llama.cpp's first stable release" — a maintainer thread from earlier this year (discussion #16111) has ggerganov stating flatly there's no stable-release notion for the project — but it's a real, verifiable step toward one.
What's actually in it
The changelog claims "100+ commits since v0.1.2," and the two most-cited additions — Granite SWA and LFM2/LFM2MOE support — aren't new to this window; they're accumulated work folded into the version bump, not something that landed since our last edition. What is new since our last check (2026-08-21T04:18 UTC): b10549 enabled tensor-split (multi-GPU) support specifically for LFM2/LFM2MOE — relevant to two of our own catalog entries in principle, useless on a single phone in practice. b10568 partially reworked Metal's RoPE handling for DeepSeek-2-family models; no before/after numbers were published, so we're not inventing one. And b10569 added a genuinely new architecture, dots3-note — except it's rednote's dots3-note-prev, a 280B-parameter, 16B-active multimodal MoE with a separate 7B vision encoder and 800M audio encoder. Server-class, not phone-class, by a wide margin.
One catalog candidate, checked and skipped
The only sub-2.5GB chat GGUF we found worth naming this window was openbmb/MiniCPM5-1B-GGUF — vendor-quantized, Apache-2.0, Q4_K_M at 688MB, comfortably under our size cap. It doesn't clear the freshness bar: it shipped roughly three months ago, well outside our 14-day window, and it only qualifies under the alternate "tops a leaderboard this week" clause if a tie for the top of Artificial Analysis's sub-4B "Tiny" category counts as topping it — we couldn't independently confirm that exact standing from the leaderboard itself, only from a secondary summary, so we're not stretching the rule on an unverified number. Everything else in Hugging Face's trending-GGUF stream this window was 27B and up. Catalog stays unchanged: 8 Generalists, 14 pre-cap specialists.
Edge in the Wild, skipped this edition
We scouted for a new DIY spotlight this run and came up short on a strong pick, not an absent one. The best lead — an ESP32-S3 project targeting zero external PSRAM, real measured numbers (20 tok/s, 161KB peak SRAM, a 24-hour soak test) — has no photos or video anywhere in its own repo, by the builder's own documentation. That's a real gap against our "real footage first" bar, and it would also be the fifth ESP32-class microcontroller project we've featured since August 1. The memory-engineering angle is genuinely distinct from the previous four, but between the missing visuals and the repetition, it's not a confident yes today. Skipping beats forcing it.
What's actually new today
One piece instead of two, and it's the more interesting of the pair this edition: a July arXiv paper measured 4-bit quantization making a vision-language model's decode 55.8-55.9% slower on real Jetson hardware. We quantized a chat model the same way, on our own engine, and measured decode 81% faster. Both numbers are real — the full breakdown, with our own reproducible benchmark, is here.