One million downloads, still no receipts

Yesterday’s tape line said it was 50k downloads from a million. This morning the repo sits at 1,049,586. Fifteen days from creation to a million for DavidAU’s NEO-CODER-MAX MTP GGUF — the flagship of his uncensored, light-thinking Qwen3.8-27B fine-tune family. What you’re downloading: the same 27B shape you already know, with the thinking tokens cut hard (his claim: one-half to one-tenth), MTP speculation folded in, and the filter off. It runs on the 24–32 GB machines most of you have.

Why the million happened is easy: it’s fast, it doesn’t lecture, it doesn’t ramble — the three things hobbyists actually click the download button for. The download button measures distribution, not judgment. What the million doesn’t tell you: there is still no independent eval of this build, and “735 ARC-C” remains his own scoreboard, not a leaderboard. Meanwhile the plain Qwen3.8-27B GGUF that everyone actually codes with still outsells it about 8-to-1 in the same 30-day window (8.86 million downloads). This is a niche’s million. A big niche — but the incumbent is still the incumbent.

So: as a coder’s upgrade, no — keep the base Qwen and buy the same effect with effort dials you can actually measure. As a fast, low-thinking, unfiltered companion for drafting and chat on one machine, fine — a million installs have road-tested the edges for you. The milestone changes the fame, not the receipts.

The machinery behind the million explains the number. Of the 126 brand-new repos that cleared 200 downloads since yesterday, 56 came from one account, 11 of them the same RP character in a trenchcoat — volume, not signal. One is a squat-shaped re-upload of the plain Qwen GGUF; another’s title is five rival model names, and if the title is the product, you’ve seen the whole product.

The big lane kept moving underneath all this. antirez’s V4.1-Flash GGUF is still compounding — 258,558 to 336,741, +78,183 in a day — and the quant wave around it holds with it. unsloth’s DeepSeek shelf still stops at the old V4-Flash: five days now of the biggest quant house in the ecosystem watching the month’s biggest model serve itself. If you have a 128 GB+ Mac or a Spark rig this is still a real deployment option, and the ds4-only lock-in is still the price.

What splitting a model across machines actually buys

meshllm/gemma-4-E4B-it-Q4_K_M-layers — a Gemma-4 build re-cut into per-layer artifacts so one model runs over a Mesh LLM cluster behind an OpenAI-compatible endpoint — pulled 44,769 downloads in two days. Know the physics before wiring up the second box. Cut by depth: machines own different layer groups and one activation vector hops A→B per token; traffic is tiny, so a slow cable is survivable, but throughput stays chained to the slowest box. Cut by width: every machine holds a slice of every layer and partial results get summed after each one — that traffic runs at token rate, it’s brutal, and it only pays with NVLink-class interconnects, which is why llama.cpp fixing its fused-QKV tensor-parallel split (b11009) is the signal that people are using it. Cut by experts (MoE only): each expert goes wherever it fits — and the single-machine version of that trick, streaming experts from SSD, was two weeks’ ago’s story.

The rule is simple: a cluster buys capacity, not speed. The counterexample is already on Apple Silicon — mlx-serve decodes 185 tok/s aggregate across 32 streams on one M4 Max. Its v26.9.3 made Qwen3.8-Flash-Next fit a 64 GB Mac: a 3.3-bit pack at 52 GB resident, claiming 85.6% top-1 agreement with bf16 (the author’s own number), and v26.9.4 added those 32 streams. On the DeepSeek side, ds4 landed batched MTP serving and Metal batching on Sep 16 — same race, turning from “fit it” to “serve many.” Before wiring two machines together, make sure one bigger isn’t better.

One more worth knowing: tencent/AuK, a 1.5B MIT speech model that unifies voice cloning, generation, content editing, enhancement, and source separation under natural-language instructions, trained on 1.95M hours (paper 2609.08936). The weights opened Sep 9 and only surfaced on HF trending today, at 2,753 downloads. The editing-and-separation shape is genuinely new; for plain TTS, Breeze still runs.

Everything else this morning is hammering the one-box ceiling — smaller packs, more streams, smarter cuts — and for anyone without an RDMA cable, that’s the correct instinct: a better single machine beats a cluster.

Sources: HuggingFace API for every download and like figure (30-day window, re-checked this morning); GitHub releases and APIs for every tooling claim; the AuK model card and paper 2609.08936. Author-reported claims — the thinking-token cut, mlx-serve’s bf16 agreement — are flagged as their authors’ own.