Open-source model ecosystem: DeepSeek, Qwen and market dynamics

Open-source projects like DeepSeek and Qwen report architectural and training gains, shifting the ecosystem to multi-model competition and experience-driven user lock-in.
Key insights
DeepSeek achieves parity with closed models via sparsity and distillation
DeepSeek v3.2 claims architectural improvements—sparse attention (DSA), expert-model distillation and RL-style training on synthetic tasks—that produce high reasoning and agent performance while lowering compute costs. The paper reports parity or superiority to contemporaneous closed models on several reasoning and agent evaluations.
Experience-fit drives user retention more than raw benchmarks
Analysis introduces the 'Cinderella Glass Slipper Effect': users who first find a model that fits a high-value workflow become locked in despite later better models. The finding implies experience match, latency, persona and ecosystem integrations matter more for retention than peak metric improvements.
Qwen releases expanded VL models and FP8 inference builds
Qwen released multiple 80B/235B model variants (Qwen3-Next series) with FP8 precision for faster inference, and Qwen3-VL-235B leads for image processing on OpenRouter. Alibaba's Qwen activity indicates continued open-source, high-capacity offerings in the marketplace.