Qwen3.8 2.4T A95B: The Open Weights Behind Qwen3.8 Max

In mid-August 2026 Qwen published the weights of a 2.4-trillion-parameter model and called it the open-source version of its Qwen3.8 flagship. That puts it among the largest open-weight models available, in the same bracket as Kimi K3 at 2.8T.

The name invites a shortcut: these are the Qwen3.8 Max weights, so Qwen3.8 Max benchmarks apply. The model card itself says otherwise, and the licence has conditions that only bite once you are successful. Both are worth understanding before you plan anything around it.

What is in the weights

The Hugging Face model card describes a causal language model with 2.4T total parameters and 95B activated per token. The details:

• 92 layers, hidden size 8,192.
• Hybrid attention: 23 blocks, each three Gated DeltaNet layers followed by one gated full-attention layer. Gated DeltaNet is a linear-attention design, so three quarters of the layers do not keep a key-value cache that grows with context.
• Mixture of experts: 512 experts, with 10 routed and 1 shared expert active per token.
• Multi-token prediction, trained with multiple steps, which serving frameworks can use for speculative decoding.
• Context: 262,144 tokens natively, extensible to 1,010,000.

For scale, Kimi K3 activates about 104B of 2.8T, and DeepSeek V4 Pro activates 49B of 1.6T. At 95B active, this model's per-token compute is in K3's bracket, roughly twice DeepSeek V4 Pro's. Active vs total parameters explains why the active figure sets your serving cost and the total sets your memory bill.

Text only, and always thinking

Two properties of the open checkpoint differ from what the hosted flagship advertises.

First, it is text only. The card states that multimodal inputs are not supported. Alibaba describes the hosted Qwen3.8 Max as having native vision; these weights do not.

Second, thinking cannot be turned off. Every response starts with a reasoning block. You can set reasoning_effort to xhigh, which is the default, medium or low, and a preserve_thinking option, on by default, keeps earlier reasoning in the conversation history.

Both defaults push cost up. Maximum effort produces the most thinking tokens, and preserved thinking makes every later turn's prompt longer. If you are serving this for routine work, set the effort explicitly and decide whether old reasoning needs to stay in context. The token cost of reasoning models covers how to measure the difference.

The scores on the card are not for these weights

The card's benchmark table is labelled as results for Qwen3.8 Max, the official hosted version, and the card notes that the hosted model has more features than the open weights. The figures are strong: 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 92.6 on GPQA Diamond, 73.5 on FrontierSWE and 92.9 on the 8-needle MRCR v2 test at 256K.

Read them as a best case for this checkpoint, not a measurement of it. The card lists what the hosted model adds: vision input, a non-thinking mode, a 1M window by default and official built-in tools. Any score that leans on those features, or on a visual benchmark, cannot have come from text-only inference on these weights.

The fix is the usual one. If the open model is a serious option, run your own evaluation on it, through the provider or the framework you would use in production. How to benchmark LLMs on your own work has a setup that fits.

The licence is permissive until you are large

The weights do not ship under Apache 2.0 or MIT. They ship under a Qwen3.8-Max licence that reads like a permissive licence with two commercial conditions. In summary, and not as legal advice:

• Broad grant. You may use, modify, fine-tune, host, distribute and sell, keeping the copyright and permission notices.
• Name display at scale. A commercial product above 100 million monthly active users or US$20 million in monthly revenue must display the model name prominently in its interface.
• Separate licence for some businesses. If you run a Model-as-a-Service or AI work-assistant business, such as a coding or office assistant, and it exceeds US$50 million in revenue over any twelve months, you need a separate licence from Qwen.
• Internal use is exempt from that second condition, as long as the model and its outputs are not offered to third parties.

For most teams neither threshold is close. For an inference provider or a coding-assistant company, the second one is exactly the business model, so read the text with counsel. Its sibling, the Qwen3.8 27B, is plain Apache 2.0, and GLM 5.2 ships under MIT. Open weights vs open source covers why the distinction matters.

What serving it takes

The weights are published in BF16. At two bytes per parameter, 2.4T parameters is roughly 4.8 TB before any key-value cache or activation memory. Qwen also publishes an FP8 checkpoint using fine-grained quantisation with a block size of 128, which it says performs nearly identically to the original; at one byte per parameter that is roughly 2.4 TB.

Either way this is a multi-node deployment. The card names SGLang, vLLM and TokenSpeed as recommended frameworks and does not list hardware requirements. Before buying or renting anything, work through model weights storage planning and the self-hosting break-even point; at this size, utilisation decides whether self-hosting is cheaper than renting.

Renting is straightforward. OpenRouter lists it at $2 input, $6 output and $0.25 cache read per million tokens at the time of writing, served by Alibaba, Novita, SiliconFlow (in FP8), Venice and Together. The providers do not agree on context: most list around 1M tokens, Venice lists 262,144. Check the endpoint you are routed to before you send long prompts.

Who should use it

• Use it if you need a frontier-class model you can host inside your own boundary, your work is text, and you can run a multi-node cluster well.
• Rent it if you want the option of self-hosting later and want to validate behaviour now on the exact weights you would run.
• Use the hosted Qwen3.8 Max if you need vision or want the model the published scores describe. The Qwen3.8 Max guide covers it.
• Use the 27B if a single GPU and an Apache licence matter more than the last few points of capability.