Qwen3.8-2.4T Raises Open-Weight Bar

Qwen is making frontier-scale capability accessible while MoE sparsity keeps inference somewhat practical—but the hardware requirements still limit true self-hosting to well-funded teams.

• –512 experts with 11 activated per token deliver a roughly 25:1 sparsity ratio
• –Native 262K context, extendable to about 1.01M tokens, suits long-running coding and research agents
• –Day-one vLLM support and FP4 quantizations improve deployment flexibility across NVIDIA and AMD hardware
• –The model requires multiple datacenter-class GPUs, so “open weights” does not mean consumer-friendly
• –Open access to a Max-tier checkpoint could accelerate distillation, fine-tuning, and independent evaluation