State of Open Models: Summer 2026 Observations

Source: HuggingFace Date: 2026-08-14 URL: https://huggingface.co/blog/state-of-open-models-summer-2026

Summary

HuggingFace’s mid-year retrospective finds Chinese labs now routinely ship the largest open-weight models of any given month (up to 2.78T parameters, vs. under 130B from US labs), that Qwen has become the de facto foundation layer with over 151,000 derivative repos (2.6x Meta’s footprint), and that actual usage is dominated by small sub-1B models (83% of downloads) running via quantized local inference — while a new, fast-growing category of “agent traffic” (Claude Code initially at 44% of agent-tagged requests) is starting to show up as a distinct usage pattern platforms can’t yet fully track.

Implications

• Open≠local / cloud-tier thread. Confirms the recurring pattern this month: frontier open releases (trillion-parameter Chinese MoE models) are cloud-tier by necessity, while the actual deployed base is small models on consumer hardware via llama.cpp-style quantization — landscape marker, not a hardware recommendation.
• Monetization-without-licensing thread. Chinese labs’ permissive Apache/MIT licensing plus API/hardware-partnership monetization matches the DeepSeek-V4-Pro pattern (MIT release, day-one Codex integration) — value capture has moved off the license and onto the ecosystem.
• Agent-traffic-as-new-signal. The first hard number on agent (not human) usage share is worth tracking release over release as a leading indicator of which harness is actually winning adoption, independent of benchmark claims.