Qwen3.8-Max ships open weights, but not under Apache 2.0
On August 12, 2026, Alibaba published the weights of Qwen3.8-Max, a 2.4-trillion-parameter MoE model, under a custom license with revenue thresholds rather than Apache 2.0. Before you deploy, read the clauses: the checkpoint is text-only and resale above a threshold becomes paid.
August 3, 2026. Qwen3.8-Max goes generally available on QwenCloud. August 12, 2026. Alibaba publishes the model’s weights as Qwen3.8-2.4T-A95B on Hugging Face. August 14, 2026. The smaller dense sibling Qwen3.8-27B follows, this time under Apache 2.0. Why it matters: for the first time, a frontier-class “open weights” model ships with a license carrying revenue thresholds — and the downloadable checkpoint is not the model the API sells.
What “open weights” does not mean
Two things that parts of the coverage blurred deserve separating. Qwen3.8-Max 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. The weights are downloadable, so it is open weights — not open source in the OSI sense, since the license is neither Apache 2.0 nor MIT.
The license attached to Qwen3.8-2.4T-A95B is a custom Qwen3.8-Max License, MIT-like in structure but with riders. The published checkpoint is text-only: vision input and the 1-million-token context stay on the hosted API. Thinking mode is forced, where the API exposes controls. If your team evaluated the model through the API expecting to self-host the same thing, you are not getting the same thing.
The license, clause by clause
The structure is an MIT with three thresholds attached, and the third one is the one that costs.
Clause 1 — Notice. Keep the copyright and permission notices in all copies. Standard, free, applies to everyone.
Clause 2 — Attribution at scale. If a commercial product exceeds 100 million monthly active users or $20 million in monthly revenue, display the model name prominently in the UI. This is a branding requirement, not a payment.
Clause 3 — The one that costs. If you or your affiliates run a model-as-a-service or AI work assistant business and aggregate revenue exceeds $50 million trailing twelve months, a separate license is required before commercial use. The word “aggregate” and the mention of affiliates read as a group-level threshold, not a product-level one.
Clause 4 — The carve-out. Internal use is exempt, provided you do not expose the model, its outputs or its capabilities to third parties. A customer-facing feature powered by the model is an exposure of its capabilities.
Who this license is actually aimed at
Clause 3 is not aimed at startups. A $50 million TTM floor with affiliate consolidation excludes almost everyone who would download a 2.4-trillion-parameter model onto a rack they own.
It is aimed at model routers, inference marketplaces and AI work assistant products that would have been the biggest winners from Max-class weights under Apache 2.0. Those players 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.
The comparison with the same week’s sibling is instructive. Qwen3.8-27B, shipped on August 14, 2026, is Apache 2.0 with no conditions, native vision and a 262,144-token context extendable to about 1 million via YaRN. The smaller model got the cleaner deal on every axis that matters to a small team, and that is no accident: you give away the model that cannot threaten the API, and you attach terms to the one that can.
The hardware contrast is just as stark. Qwen3.8-27B runs on a single 24 GB consumer GPU and posts 61.7 on SWE-Bench Pro — strong for its size, but in a different class from a frontier model. The Max, meanwhile, needs a GB300 NVL72 rack. The distance between “I can download it” and “I can serve it” is the same as the distance between the two licenses.
What the numbers say, with caution
Almost every available benchmark is vendor-published, and independent replication has not landed yet. With that caveat, the figures for the Qwen3.8-Max-0902 variant, refreshed on September 2, 2026, show a clear jump over the initial baseline. On TerminalBench 2.1, the model posts 86.6, between GPT-5.6 Sol at 88.8 and Claude Opus 5 at 84.6. On DeepSWE 1.1, it rises from 56.6 to 69.3, against 73.6 for Opus 5. On MLS-Bench-Lite, it hits 50.1, just above Opus 5’s 49.8. Separately, the AI Release Tracker listed the update at 82.7% on MMMU-Pro, a multimodal reasoning benchmark, as of its September 2 listing — further evidence the refresh was tuned for broad agent and reasoning competence rather than a single leaderboard.
The API price remains the practical argument: $2 per million input tokens and $6 per million output on QwenCloud, with a 1-million-token context. That context breaks down — per the published spec breakdowns — into roughly 991,000 input tokens, 131,000 output tokens and a separate 262,000-token ceiling reserved for reasoning steps. The split matters for agent workloads, which burn tokens on intermediate reasoning before producing a final answer.
Beside that, self-hosting a 2.4-trillion-parameter MoE with 95 billion active is not a desktop exercise: NVIDIA documented serving it on a GB300 NVL72 rack. That is data-center infrastructure. For a team that still wants to download the weights, the command is direct:
huggingface-cli download Qwen/Qwen3.8-2.4T-A95B --local-dir ./qwen3.8-maxPractically, the API is the cheaper path for almost everyone. QwenCloud at $2 in and $6 out per million tokens removes the hardware question entirely, and the hosted model is the one with vision and the full context. The weights only make sense when data cannot leave your infrastructure, or when you need to fine-tune a checkpoint you control — and only if you already own rack-scale GPUs.
What to watch
The real open question is not what Alibaba does, but whether the license structure spreads. A “open weights plus revenue thresholds” license solves one specific problem: giving away the weights to win the ecosystem while protecting API revenue from routers and inference marketplaces.
If Meta, DeepSeek or Moonshot adopt the structure on their next two releases, the “open weights” category loses its binary meaning. Legal teams will have to read every license model by model, instead of assuming a downloadable checkpoint is free to use. For now, Qwen3.8-Max is the most legible case of the trade-off: frontier weights, but a boundary that starts to cost the moment you cross it as a reseller.
Verdict
Qwen3.8-Max is a real milestone: downloadable frontier-class weights, but attached to a threshold-based license that protects the vendor’s API, and a checkpoint that is not the model being sold.
If you run internal tooling only, download the weights: the carve-out covers you outright. If you resell model access as a business, check your aggregate TTM — affiliates included — against the $50 million threshold before you deploy, and put the text in front of counsel rather than trusting a summary. If you need vision or the million-token context, use the API: the weights do not have them. And if you want unconditional open weights at this scale, they do not exist yet — Qwen3.8-27B under Apache 2.0 is strong for its size, but it is not a frontier model.