Qwen3.8-2.4T-A95B is the open-weight checkpoint behind Qwen3.8-Max, the first Max-class Qwen released with weights. Qwen describes it as built on the Qwen3.5 architecture, and the release bears that out: it is the same Qwen3_5MoeForCausalLM design scaled up rather than redesigned. Its 92 layers are 23 repeats of three Gated DeltaNet layers and one gated-attention layer, each followed by an MoE block, at width 8,192 instead of 4,096. The checkpoint holds 2,446,182,725,504 parameters, including a 26.4B MTP layer; counting the embedding and LM head, 95.3B are active per token.
The expert pool keeps Qwen3.5's shape (512 experts, 10 routed plus one gated shared expert, softmax top-k routing) but doubles expert width to 2,048, so routed experts hold 97% of the weights. Attention and DeltaNet both run at an inner width of 16,384 (2 × d) and account for about 40B of the 95B active parameters. The 23 attention layers keep 4 KV heads of dimension 256, about 92 KiB of BF16 KV cache per token (roughly 3× Qwen3.5). The 69 DeltaNet layers, now with 128 value heads, hold a fixed 552 MiB FP32 state per sequence.
The open checkpoint is text-only and thinking-only. It has no vision tower, its RoPE config drops Qwen3.5's M-RoPE sections, and reasoning cannot be disabled (reasoning_effort is xhigh, medium or low). The API model Qwen3.8-Max adds image input, a non-thinking mode and a 1M default context, and Qwen has not published how. The weights ship under a custom Qwen3.8-Max License rather than Apache-2.0.
Qwen has published no technical report. The blog describes the agentic RL system, but the pretraining data and token count, training precision, optimizer, context schedule and RL algorithm are undisclosed.