Hugging Face published its summer state-of-open-models report, covering January through July 2026 — the platform crossed one million datasets and grew from 2.43 to 2.96 million public model repositories in seven months. Before the findings, the shape underneath them, which the report states plainly: 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads. Everything below happens inside that skew, and the report's own methods note — Hub activity is one lens, not the AI market — deserves to travel with every number cited from it.
The frontier ceiling is Chinese, monthly
In almost every month of 2026, the largest open model from a Chinese lab out-sized anything American labs released of their own: China's monthly ceiling ran 754B to 2.78T parameters (Kimi K3 at the top), while America's stayed under 130B in five of seven months. The exceptions are telling by their scarcity — NVIDIA's Nemotron 3 Ultra (561B), Thinking Machines' Inkling (952B), Arcee's Trinity-Large (399B) — and the report is blunt about the rest: most US releases above 100B this year are conversions and derivatives of Chinese models, not new ones. That's the trend we called the heavyweight shift in July with a full seven months of data behind it now.
The mirror finding: US open source didn't shrink, it changed hands. The two most prolific publishers of new open models this year are AMD and NVIDIA — over 200 new repositories each, ahead of every model lab — because a model optimized for your hardware and freely available is the clearest proof the hardware works. Google and Meta, the companies that defined open model publishing, now rank well below the chip vendors. Open source moved from model labs to hardware and infrastructure companies, and the report notes the same pattern running in reverse in China, where open models are increasingly optimized for domestic chips.
Attention and adoption share exactly one model
The report's cleanest analytical move: take the top 25 repositories by 2026 downloads and the top 25 by likes. One model appears in both lists. Not a single model published in 2026 makes the download top 25; thirteen of those twenty-five date from 2022. The all-time embedding workhorse all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against 5,156 likes; Kimi-K3 collected roughly 60 downloads per like. The report's framing is worth adopting wholesale: likes measure what the field is excited about, downloads measure what it currently depends on, and treating either as a proxy for the other is the most common error in Hub coverage — "including our own earlier work," they add, which is the kind of self-correction worth crediting. The same discipline yields the small-model finding that keeps holding: among models that declare a parameter count, those under 1B take 83% of all-time downloads (61.5% of the unrestricted total — 26% of downloads go to repos with no parameter metadata), everything above 100B takes 1%, and restricting to 2026 downloads barely moves it.
The strategy split, quantified
The lab-by-lab size portfolios put numbers on a divide this blog has been describing qualitatively: Moonshot, MiniMax, Xiaomi and Z.ai publish almost nothing below 70B — frontier-only portfolios betting on benchmark position and API demand — while Tencent (59 releases) and Alibaba's Qwen (51) cover every size band from sub-1B up. The download breakdown shows how differently those bets cash out: MiniMax's 2026 volume is 100% models over 70B and Moonshot's is 88%, while BAAI's 519M downloads are 98% under 1B — same country, opposite ends of the size spectrum, both viable. The payoff gap between the strategies is stark all the same: Moonshot's frontier-only portfolio recorded 37M downloads this year; Qwen's full-spectrum family recorded 2.05 billion — roughly 55x more. And Qwen's position compounds beyond its own releases: 151,448 community derivatives, 2.6x Meta's entire footprint, growing at 180–210 new repositories per day. Of the 28,531 GGUF conversions of Qwen models on the Hub, Qwen itself published 54 — the ecosystem is community-built on top of a lab that supplied consistency, coverage, and Apache 2.0. The licensing data kills the obvious counter-story: of 178 Chinese releases above 20B this year, 81% are Apache or MIT and zero carry non-commercial terms — more permissive than the equivalent American band (29% permissive, 30% declaring nothing). Whatever the business model is, it isn't license revenue.
The quantization layer decentralized after llama.cpp joined
The monthly GGUF chart carries a structural shift the report's text mostly implies: after the ggml/llama.cpp team joined Hugging Face in February, new GGUF repositories from individual community publishers roughly doubled — 2,698 in February to 5,555 in June — while output from bulk-quantization accounts halved over the same stretch, 2,062 down to 1,297. The layer that makes trillion-parameter models locally runnable stopped being the product of a few industrial converters and became broadly distributed work, which matters because this layer is the entire reason a frontier-only release strategy reaches anyone: local inference now means a trillion-parameter MoE spread across consumer machines, and it runs disproportionately on Qwen — 39.6M GGUF downloads a month against Gemma's 20.8M and Llama's 7.5M, despite Llama-derived GGUF repos slightly outnumbering Qwen's. Same shelf space, a fifth of the traffic.
Agents are now a measurable user class
The section that couldn't have existed in March: a new dataset records the client token coding agents send when they call the Hub, making agent traffic visible for the first time. Two findings. The market has no incumbent — Claude Code held 67.8% of agent traffic in April, 6.4% in May, and 44.4% in July, swings no human-user market produces. And the fastest-growing row is "unregistered": 59.8% of agent traffic in May came from harnesses no registry had named yet, with more than a dozen new client identifiers appearing in four months. Hugging Face has spent the year building for this reader — machine-readable papers, an agents.md endpoint on every Space, an MCP filesystem tool — and the report also closes the loop on a story we covered when it broke: the July intrusion happened to them, and when their team tried to analyze the captured attack code, frontier closed models declined the work on safety grounds. The analysis was completed on a quantized open GLM-5.2 running on their own infrastructure — the gated-versus-open cyber tension this dataset has been tracking all month, compressed into one incident response.
What to expect next
- Watch whether the US frontier-scale gap closes or institutionalizes. Three original American models above 100B in seven months is a countable number; the winter report will say whether that was a trough or the new normal.
- Watch the monetization turn the report itself predicts. It expects the industry to shift toward "clearer monetization paths from open-source adoption" — the first frontier-scale Chinese release to ship with less-than-Apache terms would mark that turn precisely.
- Watch the agent-traffic share once the registry catches up. Half the story is currently in the "unregistered" row; naming those harnesses will reveal whether agent usage is concentrating or fragmenting — and the report's own bet is that the next edition "may also look very different."
References: Hugging Face — State of Open Models, Summer 2026 · related coverage: The Heavyweight Shift: Open Weights Go Trillion-Scale · Kimi K3: The Largest Open-Weight Model Yet · When the Eval Escaped: An AI Model Breached Hugging Face · GLM-5.3 · How Open Is 'Open'? · Qwen3.8-27B · Frontier Arcade: trends & predictions