Qwen Shipped 2.4T Open Weights. Apache 2.0 Didn't Survive.

TL;DR

The weights are real. Qwen3.8-2.4T-A95B went up on Hugging Face today: 2.4 trillion total parameters, 95B active per token, 92 layers, 512 experts, 213 safetensors shards, 4.89 TB on disk. It is the first Qwen-Max-class model to ship with downloadable weights, exactly as Alibaba promised nine days ago. What did not ship is Apache 2.0. Qwen3.5 and Qwen3.6 were permissive; this one carries a bespoke license named qwen3.8-max with an attribution trigger and a revenue gate aimed squarely at anyone reselling inference or shipping a coding assistant.

What actually landed

Two repositories from the official Qwen org, both updated within the same window: the BF16 original and an FP8 build using fine-grained quantization at block size 128, which the model card says performs "nearly identical" to the original. Both are text only. The API product, Qwen3.8-Max, keeps the extras: the README states plainly that 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."

So the open model is the engine block, not the car. A few specifics worth knowing before you plan around it:

• Thinking is mandatory. The card is explicit that the open model "requires thinking mode for all interactions" and that thinking cannot be disabled. There is a reasoning_effort knob (low, medium, xhigh), but there is no off switch.
• Context: 262,144 tokens natively, extensible to 1,010,000. The 1M default belongs to the hosted Max.
• Serving: the card recommends SGLang, vLLM, and TokenSpeed, with sampling at temperature 1.0, top_p 0.95, top_k 20.
• Missing: the promised Qwen3.8-27B dense companion. As of publication the official org lists only the 2.4T MoE and its FP8 twin. Community reuploads of a 27B exist; none of them are Qwen's.

The license is the actual news

Open a Qwen LICENSE file for the last two generations and you get Apache 2.0: use it, sell it, fork it, do not ask. Open this one and you get a document titled the Qwen3.8-Max License, and it has teeth in two places.

Clause 1 is attribution. If your product crosses more than 100,000,000 monthly active users or US$20,000,000 in monthly revenue, the model name "must be prominently displayed on the user interface of such product or service." Meta pioneered this move with Llama and almost nobody hits the threshold, so it mostly functions as free billboard space in the rare case someone does.

Clause 2 is the one to read twice. If you or your affiliates operate a "Model as a Service" business (offering third-party API access to inference or fine-tuning) or an "AI Work Assistant" business (AI-assisted coding or office productivity tools), and the aggregate revenue of you plus your affiliates exceeds US$50,000,000 in any consecutive twelve months, you need a separate license from Alibaba. Internal use is carved out, but only so long as you do not make the software, its outputs, or its underlying model capabilities available to any third party.

Think of it less like an open-source license and less like a rental, and more like a franchise agreement: free to cook with at home, free to feed your own staff, but the moment you put a sign out front and clear $50M, corporate wants to talk terms. That is a coherent commercial position. It is just not the position Apache 2.0 takes, and the difference matters if you were planning to build a product on the assumption that Qwen weights stay unconditionally free.

One caveat worth stating straight: the operative text is the LICENSE file itself, not this summary. The definition of "AI Work Assistant" explicitly excludes single-purpose tools and domain-specific assistants, which is a meaningful carve-out with fuzzy edges. If your revenue is anywhere near those lines, read the file and involve a lawyer rather than a blog.

"Open weights" you almost certainly cannot open

The download is 213 shards and 4.89 TB. That is not a model you pull down on a whim, and it reframes what "open" buys you here. Community quantization arrived fast, as it always does: the Unsloth GGUF repo already lists builds from BF16 down to sub-2-bit.

Read that bottom bar again. The most aggressive 1-bit quant anyone has produced is 397 GB, which means your "runs it locally" plan begins with buying a server and ends with an argument about whether a 1-bit 2.4T model still counts as the model. The FP8 build is the realistic serving path, and it still wants a cluster. For the homelab crowd, this release is a spectator sport with an unusually good view.

Where it does change things: anyone who wanted to fine-tune, distill from, audit, or red-team a frontier-class model without asking permission now can, subject to Clause 2. Weights you can inspect are worth something even when they are weights you cannot host.

What the card claims

The published table puts the open model against Opus 4.8, Fable 5, GPT-5.6 Sol (max), and its own predecessor Qwen3.7-Max. These are Alibaba's numbers, not independent runs, and no third party has reproduced them yet:

The shape matches what Alibaba published at launch: strong on research reproduction and terminal tasks, noticeably weaker on SWE-bench Pro and agentic software work where Fable 5 still leads. The useful thing about open weights is that this table is now falsifiable by anyone with enough hardware, which is a short list, but not an empty one.

Key Takeaways

• Alibaba published Qwen3.8-2.4T-A95B on Hugging Face: 2.4T parameters, 95B active, 92 layers, 512 experts, 213 shards, 4.89 TB. An FP8 build ships alongside it with near-identical reported performance.
• The license is not Apache 2.0. It is a custom Qwen3.8-Max License, breaking the permissive pattern of Qwen 3.5 and 3.6.
• Attribution kicks in above 100M MAU or $20M monthly revenue. A separate Alibaba license is required for Model-as-a-Service or AI coding/office assistant businesses above $50M in aggregate revenue over twelve months.
• This is the open engine, not the hosted product: text only, thinking mode cannot be disabled, 262K native context. Vision, non-thinking mode, 1M default context, and built-in tools stay behind the Max API.
• Practically nobody runs this at home. The smallest community quant is 397 GB; FP8 on a cluster is the realistic path.
• The promised Qwen3.8-27B dense model has not appeared from the official org yet. Anything labelled 27B on the Hub right now is a community upload.

Sources: Qwen3.8-2.4T-A95B model card, Qwen3.8-Max License, repository file listing, Qwen3.8-2.4T-A95B-FP8, Unsloth GGUF quantizations, Qwen3.5-9B (Apache 2.0), Qwen3.6-27B (Apache 2.0)