Open weights, closed features: Qwen's 2.4T Max ships text-only under a scale-tiered license — Hugging Face
Alibaba released Qwen3.8-2.4T-A95B, the first Max-class Qwen with downloadable weights: 2.4 trillion parameters, 95 billion activated, 262,144 native context. The release drops vision and video, forces thinking mode, and carries a custom license rather than Apache 2.0.
Positive-sum angle: A downloadable frontier-scale model still resets the self-hosting floor for everyone who cannot buy flagship API capacity. Regional inference hosts, quantisation communities and universities now hold a Max-class checkpoint that no vendor can withdraw from them.
What's the impact: SEA teams with sovereignty or data-residency mandates should test this checkpoint against their own workloads, then read the license tiers before pricing a service on it. Reasoning-only output changes latency budgets and cost per call.