Back to Insights
AI

Qwen3.8-Max's Weights Are Here. The License Isn't Apache and the Checkpoint Isn't the Product.

Alibaba's Qwen3.8-Max weights are live: a text-only, thinking-only checkpoint under a custom license with a $50M commercial gate. Read both before hosting.

S5 Labs Team August 12, 2026

Alibaba’s Qwen team made the flagship weights public on Hugging Face on August 12, nine days after promising them for the week of August 10. The repository is called Qwen3.8-2.4T-A95B rather than Qwen3.8-Max, and the model card explains the difference in its second callout box: Qwen3.8-Max is “the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools.” Those features live on Qwen Cloud at $2 per million input tokens and $6 output. What you can download is the text-only core underneath them.

The license is a bespoke “Qwen3.8-Max License” rather than Apache 2.0, tagged license:other, and it was committed the same morning the repo went public. Both facts matter more than the parameter count for anyone deciding whether to self-host, and neither was in Alibaba’s announcement.

What is in the repository

The commit history is short and worth reading. A folder upload landed on August 9, a chat template update on August 10, a model card on August 11, then a config fix and the LICENSE at 10:21 UTC on August 12, with the first community thread at 15:04 UTC that afternoon. The Qwen3.8-2.4T-A95B repo holds 213 BF16 safetensors shards totaling 4.89 TB; a sibling FP8 build (block size 128) is about 2.5 TB. The architecture is the one the August 3 launch described: 2.4 trillion parameters with 95 billion active, 512 experts with 10 routed plus one shared per token, 92 layers in a hybrid layout of three Gated DeltaNet blocks for every one full-attention block, and a multi-token-prediction head.

The card and config differ from the product Alibaba benchmarked and priced in a few places.

It is text-only. The pipeline tag is text-generation, the model type is qwen3_5_moe_text, and the card says “multimodal inputs are not supported.” The hosted Qwen3.8-Max takes image, text, and video input.

Thinking cannot be turned off. The chat template raises an exception if enable_thinking is set to false, and every response opens with a <think> block. What you get instead is a reasoning_effort dial (xhigh by default, medium, low) and preserve_thinking on by default. Alibaba’s own guidance is to budget 262,144 output tokens for reasoning and 131,072 for the final answer, a cost profile most self-hosters will need to tune down before production.

The native context is 262,144 tokens, “extensible up to 1,010,000” per the card, while the hosted model advertises 1M by default with a 991K input ceiling. The 262,144 figure is the config’s max_position_embeddings, not a marketing footnote. Reaching the hosted number on your own hardware means running the extension yourself and validating quality at the far end.

The benchmark table on the card is labeled Qwen3.8-Max, the hosted product name, and the footnotes describe evaluations run with a 256K context window through the Claude Code harness. The card does not say whether any of those numbers were produced with the downloadable text-only checkpoint. For the coding and agent rows the distinction is probably small; for anyone planning to use the model as a general assistant, it is not, and the card leaves it open.

Within four hours of the repo going public, a thread titled “Huge disappointment: Qwen 3.8 open weights are text-only and stripped of Qwen 3.8 Max features” was the most active on the page. The first four threads posted asked where the promised 27B sibling was, since it had not shipped alongside.

The license has two gates and no percentage

On August 7 Reuters reported, citing two people familiar with the plans, that Alibaba would require large commercial users of the open-weight Qwen3.8-Max to share revenue, with the rate still under negotiation. The LICENSE file does not deliver that. There is no percentage in it; what it contains instead is more open-ended.

The text is short. It opens with an MIT-shaped grant: use, copy, modify, sublicense, sell, deploy, host, fine-tune, “without restriction,” subject to two conditions. The first is attribution: any commercial product built on the model that passes 100 million monthly active users or $20 million in monthly revenue must display the model name prominently in its interface. The second is the one that matters:

If the licensee or any of its affiliates conducts a Model as a Service or AI Work Assistant business, and the aggregate revenue of the licensee and its affiliates exceeds US$50,000,000 (or the equivalent amount in any other currencies) during any consecutive twelve (12) months, the licensee shall obtain a separate license from Qwen before Using the Software or its derivative works for any commercial purpose.

The drafting is unusual. The revenue test is the licensee’s aggregate revenue, across all affiliates, not revenue earned from the model. The trigger is conducting a Model-as-a-Service or AI Work Assistant business at all, and once tripped, the separate license is required for “any commercial purpose,” not only for the MaaS product. And “AI Work Assistant” is defined as “an independent AI-powered product primarily designed for AI-assisted coding or office productivity,” with Alibaba’s own Qoder and QwenWork named as the examples. Single-purpose tools, assistants for other domains, and AI features inside a product whose main purpose is something else are carved out. Internal use that never exposes the model or its outputs to a third party is exempt.

So a $60 million ARR company that sells a coding agent cannot legally ship this model in any commercial product without first getting terms from Alibaba, and the LICENSE does not say what they are. That is where the revenue share Reuters described would live, if it exists: inside a negotiated agreement the public file only points to. The clause is aimed at the two categories of company Alibaba competes with, inference resellers and coding-assistant vendors, and it keys on company size rather than usage so a large competitor cannot start small.

The structure is close to a copy of Moonshot’s Kimi K3 license, which used the same 100M MAU and $20M monthly attribution test and the same “separate agreement for MaaS” mechanism. The differences are that Moonshot’s MaaS gate sat at $20 million in twelve-month revenue where Alibaba’s sits at $50 million, and that Moonshot did not name coding or office assistants as a gated category. Reuters, citing one person familiar, put Moonshot’s negotiated share at up to 30%. Alibaba’s number is not public, and until a signatory of the separate license says otherwise, “revenue share” is a reported plan rather than a term of the license.

For most teams reading this, none of it bites. A 40-person agency fine-tuning the model for a client, or a mid-market operator running it inside an internal workflow, is under both thresholds and outside both business definitions. What has changed is the default. Qwen’s open releases were Apache 2.0 for years, the same license Meta chose for Muse Glimmer 30B two days earlier, and Qwen was the lab that made the license question boring. At the Max tier it has stopped being that lab.

What it costs to run the thing you can download

The license is the cheaper problem. The FP8 checkpoint is roughly 2.5 TB of weights before any KV cache is allocated. An eight-way B300 node at 288 GB per GPU has about 2.3 TB of HBM, so the FP8 build alone does not fit on one node, and the BF16 build needs roughly seventeen of those GPUs for the weights by themselves. Nvidia’s day-zero deployment post ran the model on a GB300 NVL72, a 72-GPU rack, and reported over 4,000 tokens per second per GPU and over 350 per user in FP8, with an NVFP4 build to follow.

That is a heavier procurement problem than K3’s roughly 1.4 TB at MXFP4, and the same conclusion applies with more force. Divide the rack’s hourly cost by your sustained output tokens and compare it to $6 per million on Qwen Cloud; unless you are running it flat out, or your data cannot leave your building for reasons that override the arithmetic, the API wins. The people who benefit from these weights being public are inference providers with idle fleets and organizations whose regulatory position rules out a Chinese endpoint. The first group is exactly who the $50 million clause is aimed at.

What to read before you self-host any open-weight flagship

Alibaba did nothing outrageous here. The weights are real, they are the flagship’s core, and the terms are more permissive than any closed model will offer. What the release shows is that “open weights” from a Chinese flagship now means three separate documents to read, and the launch post is a summary of none of them.

Open the LICENSE file first and search it for revenue, users, affiliates, and separate. Note whether the thresholds are on your revenue from the model or your revenue as a company, and whether they extend to affiliates. Check whether your product category is named. If the file says “separate license” without stating terms, assume a negotiation and price the risk that the terms change after you have built on the model.

Then read the model card as a diff against the hosted product. Look for a sentence, usually near the top, that describes what the API version has that the checkpoint does not. Check the pipeline tag and model type for modality, the chat template for whether thinking is optional, and max_position_embeddings in the config for the real native context. Compare the benchmark table’s column header to the repository name; if they differ, ask which one the numbers were run on.

Finally, read the file list and multiply. Total the safetensors sizes, add KV cache at the context you intend to run, and see whether it fits hardware you can get. Kimi K3 showed that the barrier at this scale is procurement rather than permission. Qwen3.8-2.4T-A95B adds a second lesson: the free download is a text-only version of a multimodal product, with a license that reserves the right to send an invoice once you are large enough to be worth one. Plan on the checkpoint you can run and the terms you can read, and treat everything else in the launch post as a description of something you would be paying for anyway.

Sources

Want to discuss this topic?

We'd love to hear about your specific challenges and how we might help.

Start a conversation