Hugging Face hosts 2.96 million public model repositories. Yet 1.5% of them account for 99.2% of downloads, according to Hugging Face’s summer report, published August 14, 2026. That is less a long tail than a comet: one dense head, followed by an impressive amount of space.
The useful version of “Hugging Face model downloads explained” starts by refusing to turn the Hub into another fake leaderboard. Downloads can reveal operational dependence; likes reveal attention; derivatives reveal where developers keep building. None of those numbers, alone, tells you which model will solve your task, fit your hardware, or survive a licensing review.
A 2.96-million-model catalog with a very small active core
During the first seven months of 2026, public model repositories on the Hub grew from 2.43 million to 2.96 million, an increase of roughly 21.8%. Datasets rose from 711,000 to 1 million, while Spaces climbed from 1 million to 1.44 million. Publishing supply is not the problem.
Observed use is another matter. The August ecosystem analysis found that 85.6% of models had fewer than 200 lifetime downloads. At the other end, 1.5% of repositories captured 99.2% of all downloads. Repository count measures how much people can publish; download concentration shows how little of that catalog becomes common infrastructure.
That does not make the long tail worthless. A research checkpoint, narrow-language adapter, or company-specific fine-tune may succeed with 40 downloads because it was never meant for 40,000 users. The mistake is subtler: citing “nearly three million models” as evidence of nearly three million viable choices. Most developers face a much smaller operational market.
Hugging Face model downloads explained by the counting rules
A Hugging Face download is not a unique person carrying a complete model home. The Hub’s official download-counting rules count server-side requests for selected query files. Both GET and HEAD requests count. The default file may be config.json, while integrated libraries can define another file that best represents one model load.
GGUF repositories need a special rule because each quantized file is self-contained. Every GGUF file therefore counts, and cloning a repository containing several of them can register multiple downloads. Hugging Face says most interfaces fetch one GGUF, but the edge case matters when people promote small differences as market verdicts.
The public number also cannot deduplicate users, strip out CI jobs, or tell whether a file was ever used successfully. Cached loads may create no new request. API calls, private deployments, and models distributed elsewhere disappear entirely. Downloads are activity telemetry inside one ecosystem—useful, broad, and stubbornly not a customer count.
Comparison still works if you control the question. Use the same time window, repository type, and file format; compare direction as well as level; and treat sudden jumps as clues to investigate. A steady six-month climb can indicate growing integration. A three-day spike may indicate a new release, a popular demo, automated testing, or one enthusiastic cluster with excellent bandwidth.
That makes download velocity more informative than a naked all-time total, especially for newer models. Even then, it answers “how much Hub activity is this artifact receiving?” It does not answer “should my team deploy it?” Keeping those questions separate is the difference between telemetry and numerology.
Attention and adoption are different leaderboards
The report compared the 25 repositories with the most downloads accumulated in 2026 against the 25 with the most likes. Exactly one repository appeared on both lists. No model published in 2026 reached the download top 25; 13 of those 25 download leaders dated to 2022.
The cleanest example is all-MiniLM-L6-v2, an embedding model that recorded 1.55 billion downloads during the seven-month window but 5,156 likes. The report puts the distinction neatly: “A like says a release matters, and a download says something is wired into a pipeline that runs on a schedule.” One number catches applause. The other hears the server fans.
| Signal | What it can show | What it misses |
|---|---|---|
| Likes | Community attention | Repeated use |
| Downloads | Hub file-request activity | Unique users and quality |
| Derivatives | Downstream building | Production success |
| Task evaluation | Quality on a defined test | Your workload and runtime |
The ATOM adoption study uses the more defensible approach: it combines downloads, derivatives, inference share, and performance across roughly 1,500 mainline open models. We made the same argument in our analysis of why AI coding benchmarks fail as production scorecards. A single clean number is irresistible precisely because real selection is messy.
The distribution layer picks winners
Qwen shows what adoption looks like after launch week. Qwen-based models account for 151,448 derivatives on the Hub, 2.6 times Meta’s total footprint and 4.7 times the number of Llama repositories specifically. During the first seven months of 2026, developers added roughly 180 to 210 Qwen derivatives per day.
The lab did not build that distribution layer alone. Of 28,531 Qwen GGUF conversions, Qwen published only 54. Community maintainers handled the overwhelming majority of the quantization and packaging work that turns weights into something a developer can test in llama.cpp. Our guide to running Qwen3.6 27B locally exists because a model family needs those runnable artifacts before its benchmark claims become useful.
Size reinforces the point. Among models declaring a parameter count, sub-1B repositories captured 83% of all-time downloads, while models above 100B captured 1%. Restrict the window to 2026 and models above 70B still received only 3% of download volume. Frontier releases supply the spectacle; small models keep getting called from applications that have budgets, latency targets, and finite memory.
This is an ecosystem moat built from boring parts: quantizations, runtime support, examples, stable identifiers, permissive terms, and people willing to fix packaging at midnight. Labs can buy a launch. They cannot instantly buy thousands of maintained derivatives or the confidence that comes from a model already working across familiar tools.
The curated adoption sample adds a useful sanity check. By March 2026, it put Qwen at 942.1 million cumulative downloads versus Llama’s 476 million, yet different families still led different parameter bands and inference channels. Ecosystem leadership is not one throne. It is a row of workbenches, each with different tools already bolted down.
A five-signal model selection framework
Downloads deserve a column in your evaluation sheet, not the whole sheet. Use five signals, and force each one to answer a different question:
- Task quality: Does the model pass a representative evaluation set, including your failure cases?
- Recent velocity: Are downloads holding after launch, or are you staring at an all-time total accumulated by an older artifact?
- Ecosystem depth: Do maintained derivatives, quantizations, runtimes, and examples exist for your stack?
- License and provenance: Can you use the weights commercially, and can you trace the conversion back to a trusted source?
- Deployment fit: Does it meet your latency, memory, context, and cost limits on the hardware you will actually run?
That last test removes many glamorous candidates quickly. A 4-bit weight file is not the complete memory bill; KV cache, context length, runtime buffers, and multimodal projectors also want seats. Our practical VRAM sizing guide shows why even two nominally 30B models can require very different machines.
Apply the framework as a funnel. Start with five to ten plausible models from likes, search, or a task leaderboard. Remove candidates with incompatible terms or missing formats, compare recent activity and derivative maintenance, then run the remaining two or three against a fixed local test set. Record latency, memory, failure modes, and cost alongside quality. The model with fewer likes may be the one your operators can keep alive.
The order matters: use likes to discover candidates, download velocity and derivatives to estimate operational maturity, then test the survivors. Reversing that order lets the easiest number choose the shortlist before your actual requirements enter the room. Combining four mediocre proxies into one proprietary “model popularity score” does not create truth; it creates a decimal point with excellent branding.
Agents may change what a download means next
Humans are no longer the only clients browsing the shelves. Hugging Face’s agent-usage dataset says Claude Code generated 44.4% of agent-attributed Hub requests in July, while Codex rose from 10.4% in April to 20.8% in July. Those shares cover self-declared huggingface_hub traffic, not total agent use, and they move inside a zero-sum pool.
The unresolved question is this: when agents choose and fetch models automatically, will downloads reveal real dependence more clearly, or mostly reveal whichever defaults their maintainers shipped? Model ecosystems may crown durable winners less through launch-day intelligence than through the boring machinery that makes a model easy to run twice. The next State of Open Models report—and the next monthly agent-usage update—will show which interpretation is winning.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
