No new models today. One stranger’s GGUF added 71,000 downloads

First, the honest part. We queried the newest 500 HuggingFace repos plus the author feeds of every major open-weights lab: zero repos created since Sep 12 clear 200 downloads. The new uploads were noise — roughly ninety weight-merge dumps from one account, every one at zero. There is no new-model section today; the existing models are doing the moving.

antirez/deepseek-v4.1-flash-gguf — the calibrated Q2/Q4 GGUFs for DwarfStar, which we first led with on Sunday at 1,086 downloads — went from 135,900 to 207,097 overnight. About 71,000 downloads in 24 hours, on a file made by one person — more downloads than almost every base model released this month, and roughly 70% of DeepSeek’s own repo (288,414). Nobody pulls 340 GiB by accident. The call from Sunday stands: run it on a 128 GB+ Mac or Spark-class rig, DwarfStar lock-in included, speeds author-reported. And unsloth, for the third straight day, still has no V4.1-Flash GGUF — the biggest quant house in the ecosystem, watching a 288,000-download model serve itself.

Meanwhile, five strangers placed a format bet. Since Sep 12, at least five new “NVFP4” builds have crossed the 200-download line on four different hot bases — a hybrid re-quant of DavidAU’s NEO-CODER-MAX among them, plus a rider on a base nobody had covered until the quant arrived. Somebody also re-uploaded NEO-CODER-MAX verbatim under a lab-sounding org name and picked up 235 downloads; being cloned for traffic isn’t flattery, it’s an index of where attention lives. The riders’ numbers are small; the pattern isn’t. And at the head of the hill, NVIDIA’s Qwen3.8-Flash-Next NVFP4 checkpoint sits at 110,548 — from 1,129 on Sep 6 to 78,737 by Sep 12 — one of the fastest curves we’ve logged. Nobody has benchmarked any of it.

Block-scaled FP4: sixteen seats, one shared scale

A 4-bit float sounds like a rounding joke. One sign bit, a 2-bit exponent, a 1-bit mantissa (E2M1), and a single FP4 number can only be one of eight magnitudes: 0, ½, 1, 1.5, 2, 3, 4, or 6, plus a sign. That’s the whole vocabulary — and a transformer weight is usually something like 0.0137, which is unwriteable with these tiles. The rescue is block scaling: every small block of neighboring weights — 16 for NVIDIA’s NVFP4, 32 for the OCP standard’s MXFP4 — stores its E2M1 tiles plus one shared multiplier. NVFP4’s multiplier is an 8-bit float (E4M3) with a second, whole-tensor FP32 scale; MXFP4 keeps its multiplier a pure power of two. The arithmetic: 16 weights × 4 bits + an 8-bit scale = 4.5 effective bits per weight for NVFP4, 4.25 for MXFP4. You pay half a bit per weight to rent one shared fact: what order of magnitude this block lives in.

Why a float grid and not 4-bit integers, which quantization has had all along (your Q4_K_M is exactly that)? The shape of the grid. An integer ladder’s uniform spacing lets the block’s biggest magnitude dictate the range, grinding every small weight against the same coarse rungs. A float grid has relative spacing: dense near zero, coarser as magnitude grows — 0.5 apart down at 1, 2 apart up at 6. Trained weights cluster near zero, so FP4 spends its sixteen seats exactly where the probability mass lives. That doesn’t make outliers free — one big weight in a block of small ones still wrecks its neighbors’ scale — but FP4’s error scales with the value and INT4’s doesn’t. Calibration — imatrix-style, or rotations that shuffle outliers around — still decides whether your quantized model thinks straight. Block scaling is the stage, not the rehearsal.

Why this year. Formats don’t win on elegance; they win when the silicon starts multiplying in them natively. Blackwell-class tensor cores compute FP4 directly (NVIDIA quotes double the FP8 throughput — their number, flagged), turning weights and activations — W4A4 — from a memory trick into a compute win: fewer bytes moved and faster matmuls, in stacks like SGLang and TensorRT-LLM. That’s why NVIDIA ships first-party NVFP4 checkpoints instead of feeding the quant cottage industry.

And the catch your machine decides. On a Mac, FP4 is a capacity format, not a speed one: MLX and its kin dequantize to bf16 before the math. On a pre-Blackwell GPU, “NVFP4” buys memory headroom, not tokens. And most “NVFP4”-labeled GGUFs — hybrids included — are storage shapes: the label tells you how the weights are squeezed on disk, not that anything on your box multiplies in them. Until someone publishes that NVFP4-vs-Q4_K-vs-Q6_K sweep, “FP4 won” is a statement about Blackwell serving economics, not about tonight’s chat.

So, the practical rule. Blackwell box serving real traffic: NVFP4 through a native engine — that’s what the 110,548 downloads are. Apple Silicon: MLX 4/8-bit or an imatrix GGUF plus a drafter; ignore the format war. Everyone else: your Q4/Q6 K-quants keep the quality-per-GB throne, and the correct response to an “NVFP4 GGUF” is to check what the engine does with it, not what the filename promises.

One line from the harness lane: llama.cpp shipped v0.4.1 with official Ubuntu CUDA binaries — the “compile from source or die” era is ending on Linux — and oh-my-pi learned to recover the serving model id from the signed thinking block, a client-side receipt against routers that quietly swap your model. The rest was plumbing: the kind that decides whether an agent you leave running tonight is alive tomorrow.

Sources: every download/like figure pulled from the HuggingFace API on 15 Sep 2026, including the zero-day sweep; llama.cpp and oh-my-pi release notes and the GitHub API for harness items. NVIDIA’s FP4 rate is NVIDIA’s own, vendor-reported; the block-scaling specs (16-wide E4M3-scaled NVFP4, 32-wide MXFP4, E2M1 grid) are published formats.