Alibaba Cloud Presenta Qwen3.8-2.4T-A95B y Qwen3.8-Max: MoE Abierto de 2.4 Billones de Parámetros

Alibaba Cloud's Qwen team has shocked the global AI landscape by releasing the post-trained weights for Qwen3.8-2.4T-A95B alongside the cloud-managed Qwen3.8-Max service.

Gated DeltaNet: Sub-Quadratic Linear Attention

At 2.4 trillion total parameters with 95 billion activated per token across 92 layers, Qwen3.8 introduces a hybrid layout:

• 23 repeating blocks consisting of: 3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE).
• 128 linear attention heads for Values and 16 for Query-Key projections.
• Sub-quadratic memory scaling across ultra-long prompts.

Controllable Reasoning: reasoning_effort & preserve_thinking

Addressing the needs of complex coding agents, Qwen3.8 introduces deep reasoning control:

• reasoning_effort: Allows developers to dynamically allocate thinking tokens between 0 (instant zero-shot generation) and 64,000 thinking tokens for formal verification.
• preserve_thinking: Carries previous chains of thought across multi-turn developer sessions, avoiding token re-computation when debugging complex software repositories.