Models & Weights

Qwen3.8-2.4T-A95B

Qwen's 2.4T flagship weights land — under a license with a $50M trip-wire the API never mentions

Open code / restricted weightsModel

What it is

The open-weights release of the Qwen3.8-Max generation: a 2.4T-total / 95B-active MoE with 92 layers, 512 experts (10 routed plus 1 shared), the Gated DeltaNet–Gated Attention hybrid of its 27B sibling, and 262K native context extensible to about 1M. It is not the API model in a trench coat: this checkpoint is text-only, runs in always-on thinking mode, and ships without the vision input, non-thinking mode, default 1M context, and built-in tools the hosted Qwen3.8-Max offers.

Why it's interesting

This resolves our watch-list question from the 27B release — 'will Max's weights actually land, and under what license?' They landed August 12, and the license is the story: not Apache like the 27B, but a bespoke Qwen3.8-Max License. The grant itself is broad, MIT-flavored text — use, modify, sell, host, fine-tune. Then the conditions: past 100M MAU or $20M monthly revenue you must display the model name in your UI, and any model-as-a-service or 'AI work assistant' business over $50M trailing-twelve-month revenue needs a separate paid license from Qwen before commercial use.

Use cases

  • Self-hosted frontier-class inference for orgs comfortably under the thresholds
  • Fine-tuning and RL research on a true Max-class base rather than a distillation
  • Internal deployments — the license's cleanest lane, since purely internal use avoids the MaaS trigger entirely

Who it's for

Infra-rich teams under the revenue thresholds — after a license read with counsel

Setup

Advanced. 2.4T parameters in BF16 means a serious multi-node deployment; community quantizations were already appearing at check time

Limitations & cautions

Text-only with thinking mode you cannot turn off — the API keeps vision, non-thinking, default 1M context, and built-in tools for itself; the hardware bill is Kimi K3-class; and the bespoke license means legal review before any commercial deployment, with a revenue definition that sweeps in affiliates.

Editorial takeaway

Last week the 27B shipped Apache with a LICENSE file and we praised it. This is the other shoe: the closer a checkpoint sits to the money, the further the license drifts from open. Read the definitions — 'AI Work Assistant' may describe your product.

Related & alternatives