trending models on hugging face: qwen 3.8 lost 622k downloads this week, its 9x-compressed fork gained 480k

1. laya (convaiinnovations) – 0.4b text classification, 2.23k likes, zero downloads logged. four days old, likes-driven trend, not usage
2. ternary-bonsai-2-27b (prism-ml) – 27b, 2.57m downloads, 1.82k likes. qwen 3.8 compressed 9x to 5.9 gb via ternary weights
3. qwen-image-2.1 (qwen) – 7b text-to-image, 16.2k downloads, 1.66k likes
4. xing4.0-29b (xingchen-agi) – 31b text generation, 30.6k downloads, 1.25k likes
5. qwen-image-2.1-gguf (abenzerps) – 7b quantized fork of #3, 182k downloads, 884 likes

ternary bonsai 2 pulled 2.57m downloads in 5 days – a 27b model compressed to 5.9 gb that retains 98.2% of qwen 3.8's benchmark scores. builders are choosing the 9x smaller version over the original.

qwen 3.8 itself – the full 27b original – dropped from rank 5 to rank 7 and shed 622k downloads week-over-week. a compressed version from ista-daslab gained 480k in the same window. builders are trading the full model for the lighter one.

deepseek v4.1 flash sits at #6 with 542k downloads – the same model topping openrouter's usage charts. there's the split in one sentence: deepseek flash gets rented through an api by the token, bonsai gets downloaded and run locally.

anthropic shipped opus 5.5 today – 40% cheaper to run than opus 5, leading every agentic coding benchmark. meanwhile on hugging face, builders are pulling 11.9m downloads across the top 10 trending models that cost nothing.