Nobody downloads a 340 GiB file by accident

antirez/deepseek-v4.1-flash-gguf — the calibrated Q2/Q4 weights for DwarfStar — went from 1,086 downloads yesterday morning to 135,900 today. A 125× jump in a day, and the number is worth trusting precisely because of what it is: a 340 GiB file for one niche engine. There is no accidental clicking at that size. People are deciding to spend an afternoon and a shelf of disk to run this model, which makes it the hardest adoption signal the whole DeepSeek-V4.1-Flash ecosystem has produced. The base model is riding the same wave: 140,636 → 244,457 downloads in the same 24 hours, +74%.

The engine caught up with the appetite while people were downloading. V4.1 Flash support landed on DwarfStar’s main branch Sep 12, and antirez’s Sep 13 post dropped the number that matters: the model running on a single NVIDIA DGX Spark — the size of an espresso machine — at ~9 tokens/s over SSD streaming, or ~22 tok/s across two Sparks linked by RDMA. Author-reported from his posts, his conditions, not a published benchmark — but the hardware floor is unmistakable. Two days ago the entry ticket to this model was “a 128 GB Mac club membership.” Today it’s a small black box that costs less than a phone-flagship laptop, at interactive-if-patient speeds. His economics take: two 128 GB laptops plus an RDMA cable are out-competing a 256 GB Mac Studio at its own price — “it is just a matter of developing good inference software.” If you live in llama.cpp/Ollama on 16–32 GB, though, none of this is yours yet.

A lab that disowns its own benchmark table

Today’s only new model past our gate is Agnes-3.0-Flash Preview (Agnes AI, Apache-2.0, 33B, 262K context, text/image/video), and it opens with the paragraph most labs would never write: the production API’s benchmark results “should not be attributed to the Preview weights released here.” The benchmark table, the card admits, is “compiled from different sources, harnesses, and model snapshots” — not a head-to-head. That’s the most honest release note of the week, and I mean it as praise, not sarcasm. What’s also honest: this is a small lab whose two prior releases pulled 318 and 182 downloads, and its own quants already out-pull it — quimmedes’ GGUF at 7,416 downloads in a day vs the base repo’s 474. Kick the multimodal around if you like; the review happens when production weights ship open and someone runs a controlled eval.

People don’t download papers, they download boxes

The same tell shows up twice more. YuE2-3B, the open music generator called Marginal on Sep 12 at 971 downloads, now has its base at 3,707 — and its GGUF build at 14,077. The run-ready package out-pulls the model roughly 4-to-1, ComfyUI packaging at another 7,587. Same shape at the fine-tune layer: agentionai’s Signal-3.8-27B has 8,152 GGUF downloads against 55 for its safetensors — a 148× gap. Fine-tunes are increasingly discovered at the quant layer; the weights repo is a formality. Note for the music fans: YuE2 is CC-BY-NC — personal and research use only.

The first diffusion model that designs your UI — and it’s billed per step, not per token

OUI-1 (Thesys, Apache-2.0, 1,447 downloads) is the first diffusion-LLM fine-tune past our gate: a merged LoRA on Google’s DiffusionGemma — a June release that has quietly collected 744,625 downloads — taught to write interface screens in openui-lang. On the lab’s own Generative UI Benchmark it scores 71.7% against the base model’s 13.0%, 5.5× its own base. Vendor bench — but the raw outputs are committed to the repo so the number can be rescored offline, which is more than most release notes invite.

The interesting part is what diffusion does to the cost model. An autoregressive model is billed per token: output length is the invoice. A diffusion LM takes a 256-token canvas of noise and denoises the whole block in parallel over a fixed number of steps (OUI-1 ships the dials: 48 steps, an entropy bound of 0.1 below which tokens commit). A 2,000-token screen and a 200-token screen cost nearly the same — which is why a light UI screen lands in ~1 second on a single A100, dense ones in 3–6. The catches are real: starve the steps and structure collapses (32 steps scores like 48; 16 breaks the output entirely), text arrives one canvas at a time so chat streaming dies, and temperature and seed are ignored. Practical rule: chat stays autoregressive plus speculative decoding; if you generate whole structured artifacts — UIs, configs, function bodies — the diffusion path is now a live second answer to latency.

Loose ends, kept short

MiniCPM5-2B keeps compounding (150,110 downloads, official GGUF at 99,716) and just got a first-party speculative-decoding drafter. On DwarfStar, PR #991 is up for Qwen3.8-Flash-Next support, and a 2023-vintage repo called PI-Desktop caught fire (+550 stars/day to 3,449) because its session import keeps collapsing the switching cost between coding agents.

One thing to hold onto: the common thread isn’t a smaller frontier. It’s software making boxes people already own do things they couldn’t do last month — and the downloads now vote on it, in 340 GiB increments.

Sources: HuggingFace model cards and the HF API (figures verified 14 Sep 2026), the GitHub API for PI-Desktop’s stars, antirez’s public X posts for the Spark and RDMA figures (author-reported, not a published benchmark), and OUI-1’s and Agnes-3.0-Flash’s model cards. Vendor- and author-reported benchmarks are flagged as such.