Qwen3.8-Max

How it is built

The open weights carry the name Qwen3.8-2.4T-A95B, which states both sizes: 2.4T parameters in total, 95B active for each token.¹¹1Qwen, Qwen3.8-2.4T-A95B model card, August 2026. Terminal Bench 2.1 with Claude Code (avg@10), a 5-hour timeout and max_tokens=131,072. SWE-bench Pro with the Claude Code harness, temperature 1.0, top_p 0.95 and a 256K context. reasoning_effort levels xhigh, medium and low. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B

Its 92 layers come in 23 groups of four, three Gated DeltaNet layers followed by one Gated Attention layer.¹¹ Every layer ends in a mixture of 512 experts, and a token is routed to 10 of them plus 1 shared expert.¹¹

The DeltaNet layers have 128 value heads and 16 query-key heads. In Gated Attention, 64 query heads share 4 key-value heads.¹¹ The hidden size is 8,192, and training included multi-token prediction.¹¹

What is new

Qwen had never released the weights of a Max model before.¹¹ The API version was announced on 3 August 2026. The weights followed on 12 August.²²2Qwen, Qwen3.8-Max blog post, 3 August 2026. https://qwen.ai/blog?id=qwen3.8 ³³3Qwen, Qwen3.8 repository README, news entry of 12 August 2026. https://github.com/QwenLM/Qwen3.8 The architecture builds on Qwen3.5.¹¹

On Terminal Bench 2.1 Qwen reports 86.6, against 74.5 for Qwen3.7-Max, averaged over 10 attempts in the Claude Code harness.¹¹ On SWE-bench Pro it reports 67.7, where Qwen3.7-Max had 60.6.¹¹

Running it

The released weights read text only. They always think before they answer, and thinking cannot be turned off.¹¹ The API version adds image and video input and a non-thinking mode, with a 1M context by default.¹¹⁴⁴4QwenCloud, Qwen3.8-Max model page, prices per million tokens as listed on 29 September 2026. Input text, image and video; context 1M tokens. https://www.qwencloud.com/models/qwen3.8-max

reasoning_effort sets how deep the model thinks. The default is xhigh, the highest of three levels.¹¹ For agentic work the card suggests 262,144 tokens for reasoning and 131,072 for the final answer.¹¹

An FP8 checkpoint is published too. At one byte per parameter its weights come to about 2.4 TB.⁵⁵5Qwen, Qwen3.8-2.4T-A95B-FP8 model page. Arithmetic: 2.4T x 1 byte = 2.4 TB, before the KV cache and any tensors kept at higher precision. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B-FP8

The licence allows commercial use on two conditions.⁶⁶6Qwen, Qwen3.8-Max License, conditions 1 and 2. Internal use that exposes nothing to third parties is exempt from condition 2. https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/LICENSE A product with more than 100 million monthly active users, or US$20 million of monthly revenue, must display the model name. A company selling models as a service or AI work assistants needs a separate licence once its revenue passes US$50 million in twelve months.⁶⁶

On QwenCloud, qwen3.8-max accepts up to 991K input tokens and returns up to 131K.⁴⁴ Input costs $2 per million tokens, or $0.25 from the implicit cache. Output costs $6.⁴⁴