MON, AUGUST 17, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Large Language Models

Alibaba Open-Sourced a 2.4T Frontier Model, Then Attached Three Revenue Thresholds to It

The Qwen 3.8-Max weights went live August 12 as Qwen/Qwen3.8-2.4T-A95B, the first downloadable Max-class Qwen. They ship under a custom license rather than Apache 2.0, as a text-only checkpoint with forced thinking, missing the vision input and 1M context the hosted API still sells. Internal use is free, 100M MAU triggers attribution, and reselling above 50M dollars TTM requires a paid license.

By AIToolsRecap August 16, 2026 9 min read 43 views
Home Articles Large Language Models Qwen 3.8-Max Open Weights vs the API: What You ...
THE ANSWER, UP FRONT

● Free and unconditional — internal use where you never expose the model, its outputs, or its capabilities to third parties.

● Free with attribution — product above 100M MAU or 20M dollars monthly revenue must display the model name in the UI.

● Requires a paid license — model-as-a-service or AI work assistant businesses above 50M dollars TTM revenue, affiliates included.

● The catch nobody mentions: the weights are text-only. Vision and the 1M context stayed in the API.

What actually shipped on August 12

Alibaba published Qwen/Qwen3.8-2.4T-A95B on Hugging Face: BF16 safetensors in standard Transformers format, with an FP8 variant alongside, documented as compatible with vLLM, SGLang and TokenSpeed. NVIDIA's deployment engineering blog confirmed the release the same day in a walkthrough of serving the model on a GB300 NVL72 rack.

That is a real, confirmable release. It closes a promise Alibaba made on August 3, when Qwen3.8-Max went generally available and the team said the open weights would follow the next week.

What arrived nine days later is narrower than what was promised.

Weights versus API: the split nobody put in the announcement

Capability Downloadable weights Hosted qwen3.8-max API
Text input Yes Yes
Image and video input No Yes
Context window Reduced, 1M not carried over 1M in, 131,072 max output
Thinking mode Forced on Product controls available
Architecture 2.4T MoE, 95B active per token 2.4T MoE, 95B active per token
Price Your GPU bill 2 dollars in / 6 dollars out per million
License Custom Qwen3.8-Max License Standard API terms

The model card describes the hosted qwen3.8-max as the official version based on Qwen3.8-2.4T-A95B with more features. That phrasing is doing a lot of work. If your team evaluated the model through the hosted product and planned to self-host the thing you tested, you are not getting the thing you tested.

The license, clause by clause

This is a custom Qwen3.8-Max License, copyright 2026 Qwen. It is not Apache 2.0 and not MIT. Structurally it is MIT-style with riders attached.

CLAUSE 1 — NOTICE

Keep the copyright notice and permission notice in all copies or substantial portions. Standard, applies to everyone, costs you nothing.
CLAUSE 2 — ATTRIBUTION AT SCALE

If your commercial product or service exceeds 100 million monthly active users or 20 million dollars in monthly revenue, you must prominently display the model name in your UI. This is a branding requirement, not a payment.
CLAUSE 3 — THE ONE THAT COSTS MONEY

If you or your affiliates run a model-as-a-service or AI work assistant business and aggregate revenue exceeds 50 million dollars trailing twelve months, you must obtain a separate license from Qwen before commercial use. Note the word aggregate, and note that affiliates count.
CLAUSE 4 — THE CARVE-OUT

Internal use is exempt, provided you do not expose the model, its outputs, or its underlying capabilities to third parties. Read that as written: shipping a customer-facing feature powered by the model is exposing its capabilities.

Who this license is actually aimed at

Clause 3 is not aimed at startups. A 50 million dollar TTM revenue floor with an affiliate rollup excludes almost everyone who would download a 2.4-trillion-parameter model to a rack they own.

It is aimed at the model routers, the inference marketplaces, and the AI work assistant products that would otherwise have been the single biggest beneficiaries of Max-class open weights. Those businesses take an open checkpoint, serve it at a margin, and compete directly with the vendor's own API. Apache 2.0 permits that unconditionally. This license does not.

That creates a middle category that did not really exist before: weights you can download, inspect, fine-tune and deploy, but not resell above a revenue line. Whether Meta, DeepSeek or Moonshot copy the structure is the thing to watch over the next two releases.

Compare it to the sibling release

Qwen 3.8-Max weights Qwen 3.8-27B
Released August 12, 2026 August 14, 2026
License Custom, revenue thresholds Apache 2.0, no conditions
Multimodal No, text-only checkpoint Yes, native vision
Context Reduced from the API 1M 262,144 native, 1M via YaRN
Hardware to run it Rack-scale, GB300 NVL72 class One 24GB consumer GPU
SWE-Bench Pro Not published for the checkpoint 61.7

The smaller model got the cleaner deal on every axis that matters to a small team. That is not an accident, and it is a reasonable trade on Alibaba's side: give away the model that cannot threaten your API business, and attach terms to the one that can.

Decision framework

If you are... Do this
Running internal tooling only Download the weights. The carve-out covers you outright
Building a product under 20M dollars monthly revenue Fine, keep the notice, no attribution rider yet
Reselling model access as a business Check your aggregate TTM against 50M dollars before you deploy, affiliates included
Needing vision or the 1M context Use the API. The weights do not have it
Wanting genuinely unconditional open weights Qwen 3.8-27B under Apache 2.0, or Muse Glimmer 30B
Without rack-scale GPUs Not a real option. 95B active per token on a 2.4T checkpoint is not desktop hardware

What the model does, for context

Qwen3.8-Max went generally available August 3, 2026, following a benchmark-free preview at WAIC Shanghai in July. It is a sparse mixture-of-experts model, 2.4 trillion total parameters with 95 billion active per token, built on the Qwen 3.5 architectural foundation with hybrid attention.

On vendor-published numbers it posts 86.1 on OSWorld-Verified and leads PaperBench at 93.0. API pricing on QwenCloud is 2 dollars per million input and 6 dollars per million output with a 1M context, which undercuts Kimi K3 at 3 dollars and 15 dollars. The 95B active figure is what drives serving cost and puts it in a lighter class per token than K3's 104B active.

As with the 27B release, every launch benchmark is vendor-run. Independent replication has not landed.

FAQ

Is Qwen 3.8-Max open source?

The weights are downloadable, which is open weights, not open source. The license is custom rather than an OSI-approved one, and it carries revenue-based conditions that Apache 2.0 and MIT do not. Calling it open source is imprecise in a way that matters if you are doing legal review.

What hardware do I need to self-host it?

Rack scale. NVIDIA documented serving it on a GB300 NVL72. A 2.4T MoE with 95B active per token is not a single-server workload, let alone a desktop one. If self-hosting frontier weights is the goal but the hardware is not there, the API at 2 dollars and 6 dollars per million is the practical path.

Do the revenue thresholds apply to my whole company or just the product?

The model-as-a-service clause specifies aggregate revenue including affiliates, which reads as company-wide rather than product-level. The attribution rider is written against the commercial product or service. If you are near either line, get the text in front of counsel rather than relying on a summary.

Why would Alibaba release weights and then restrict resale?

Because the businesses most likely to serve a downloaded frontier checkpoint at a margin are the ones competing directly with QwenCloud. Releasing the weights buys developer goodwill and ecosystem adoption. The 50 million dollar clause protects the API revenue from the handful of players large enough to threaten it.

Should I use the weights or the API?

API, unless you have rack-scale GPUs and a specific reason to control deployment, such as data that cannot leave your infrastructure. The weights lose multimodal input and the 1M context, so for most workloads the hosted version is both cheaper in practice and more capable.

Is there a fully unrestricted alternative at frontier scale?

Not at Max scale right now. Qwen 3.8-27B is Apache 2.0 with no conditions and scores 61.7 on SWE-Bench Pro, which is strong for its size but is not a 2.4T frontier model. That gap is the point of the license.

Tags
QwenQwen 3.8-MaxAlibabaOpen Source AIAI GuideKimi K3Qwen 3.8-27BHugging FaceNvidia2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →