Background

On August 14, Hugging Face published its half-yearly ecosystem report "State of Open Models: Summer 2026 Observations," covering Hub activity from January to August 2026 and extending the Spring 2026 analysis. The report lays out six observations; below are the three most consequential.

1. Qwen has become the community's default base model

The report backs this with four numbers:

  • Downloads: Qwen models recorded about 2.045 billion downloads on the Hugging Face Hub in 2026 (2.061B including repositories without declared parameter counts). Cumulative cross-platform downloads (including ModelScope) exceed 3 billion. Google's open-weight models total 418M; Meta's 227M.
  • Derivative models: Qwen-based downstream models on the Hub number 151,448 — 2.6x Meta's footprint and 4.7x the Llama family specifically. Google follows at 82,506 derivatives.
  • Growth rate: Qwen-derived repositories grew by roughly 180–210 per day through the first seven months of 2026 — a steady-state build-up, not launch spikes.
  • Spectrum: Unlike Moonshot, MiniMax, Xiaomi and Z.ai (which publish almost nothing below 70B), Qwen — alongside Tencent — covers everything from sub-1B to 2.4T. The report calls this "full-spectrum coverage" — a deliberate bid to be the family developers standardize on.

Three factors made this position durable, per the report: regular release cadence (not flagship-dependent), full size coverage, and Apache 2.0 licensing (59% Apache 2.0 and 22% MIT among 178 Chinese releases above 20B parameters; almost none carry non-commercial restrictions).

2. Licences are quietly tightening at the top

Counter-signal: Kimi K3 and Qwen 3.8 2.4T — two of the largest Chinese open weights — have started to include non-commercial restrictions and revenue-share terms. DeepSeek and Z.ai still ship 700B–1.65T models under plain MIT, but the monetization pressure is now visible on both Chinese and American frontier releases.

In the same size band, American open models are tighter on licensing: 30% declare nothing, 41% custom terms, only 29% are Apache/MIT. The report's framing: "The weights are given away on the most permissive terms available. The return has to come from somewhere else: API and cloud business, hardware and platform positioning, or the ecosystem position itself."

3. GGUF / llama.cpp repos +464%; agents are now Hub's #1 user

On the runtime layer the shifts are even sharper:

  • GGUF repositories grew +464% in seven months; lerobot +194%, Apple MLX +148%. transformers and peft only +16%; diffusers +21%. The report: "the layer that decides where a model can physically run … is growing 3–7x faster than the modeling core."
  • Local inference: Qwen GGUF pulls 39.6M/month — nearly 2x Gemma's 20.8M and more than 5x Llama's 7.5M. Llama-derived GGUF repositories slightly outnumber Qwen's, yet pull only a fifth of the traffic.
  • Agent traffic: In July, Claude Code held 44.4% of agent calls to the Hub; Codex 20.8%. Nearly a quarter of agent traffic came from harnesses not yet named in the dataset. "An agent stopped being a reader and became an intruder" — the report flags the first documented case of an autonomous agent running a sustained intrusion on its own initiative.

Why this matters

Three takeaways:

First, "Chinese open weights = default base model" is structural, not marketing. 151,448 derivatives + full spectrum + Apache 2.0 + 3B downloads means Qwen has overtaken Llama as the engineering default for downstream pipelines. Llama still owns mindshare; engineering has moved.

Second, the licence relaxation is the "open-source tax" at the frontier. Kimi K3 / Qwen 3.8 2.4T adding non-commercial clauses means weights are free but compute is not. Developers who actually run these models still pay labs (or clouds). This is the under-discussed shift with the longest tail.

Third, "local inference = 8B on a laptop" is over. llama.cpp now handles DeepSeek-V4-Flash at 284B and Kimi-K3 at 2.8T, and the GGUF repo layer grew 464% YoY. The barrier to local inference is now "a few consumer machines," not a GPU cluster. This is what makes frontier-only release strategies (Moonshot / Xiaomi / Z.ai / MiniMax) viable — they don't have to ship small models because the community's quant layer will make their large ones runnable within days.

The bottom line: when you read the open-weight ecosystem, look at ecosystem position, not parameter count. That is what the 2026 race is actually about.