IT EN
Start a trial

Operational notes Observatory

Qwen3.8-Max announces open weights: the word "licence" never appears

6 min read

Two open, empty cardboard boxes photographed close-up against a plain background, in black and white
The box is open, not the contents: an announcement of openness is not yet a delivered parcel.

On 3 August 2026 QwenTeam published its announcement of Qwen3.8-Max on its own blog: “Today, we are officially releasing Qwen 3.8-Max, the most capable model in the Qwen family to date. This also marks the first time we will open-source the weights of a Qwen-Max-class model — the open weights will be released next week.” A model with 2.4 trillion parameters (95 billion active), the first Max-class model Alibaba has promised to open. 5,068 words of post, benchmarks, use cases, line-by-line integration instructions. One word never appears: “licence”. Not “license” either, not “Apache”, not “MIT”. We searched for them one by one: zero occurrences. That is where this piece starts, because for anyone who has to decide whether to run this model — defence, government, large enterprise — that missing word is not a stylistic detail.

The facts, in order

  • 3 August 2026 — QwenTeam states “2.4T parameters (95B active), with open weights releasing next week” and confirms: “Qwen3.8-Max is now available through QwenCloud (…) The model weights will be open-sourced on Hugging Face and ModelScope next week — stay tuned.”
  • Verified by us the same day — no Qwen3.8-Max repository appears under the Qwen organisation on Hugging Face; the public API endpoint (api/models/Qwen/Qwen3.8-Max) returns a 401; searching “Qwen3.8” surfaces only two third-party distillations, not the official model. The weights, today, do not exist as a downloadable file.
  • The word that’s missing — across 5,068 words, “licence”/“license” never appears, nor does “Apache” or “MIT”. The only access available today is the hosted API on Alibaba’s infrastructure.
  • The comparison, not the inference — previous open Qwen models sit under an Apache 2.0 licence, confirmed by the public pages for Qwen/Qwen3-235B-A22B and Qwen/Qwen3-ASR-1.7B-hf. That doesn’t mean Qwen3.8-Max will inherit the same text: it only means that, until it’s stated, you don’t know.
  • Secondary sources diverge, the primary one settles it — the South China Morning Post reports the model is being made “widely accessible ahead of open-weights release”; MarkTechPost writes the same day that Alibaba “has not disclosed the activated-parameter count” — but the primary source states it, 95 billion. When two secondary sources contradict each other, you go back to the original document.

Three clauses that don’t exist yet

The weights aren’t a file — for now they’re a promise with a date attached. Until they appear on Hugging Face or ModelScope, the only way to use Qwen3.8-Max is the API on QwenCloud — outside the perimeter of whoever uses it, inside the infrastructure of whoever hosts it. “Next week” is a date, not an artefact: before writing it into a specification, you need to see it, not take it on trust.

“Open-source the weights” is not a licence. The sentence that opens the announcement describes an intention, not a legal regime: it doesn’t say whether commercial use is permitted without thresholds, whether there are user or revenue limits, whether distillation is allowed, whether attribution is required. We have already measured this line by line on a text that genuinely exists: Kimi K3’s licence arrived the same day as its weights, under five hundred words, and even there a group-revenue threshold changed the substance for anyone reading quickly. The clauses that matter are never the same from one model to the next: revenue thresholds, field-of-use restrictions, revocation conditions. Until Qwen publishes a text, every assumption about Qwen3.8-Max is a guess, not a reading.

“Opening the weights” doesn’t specify what gets opened. Just the final weights, or also the post-training recipe, the data, intermediate checkpoints, official quantisations? The line between “the model is open” and “what actually stays installable” is the question that matters, and the announcement doesn’t touch it anywhere in its 5,068 words.

The base-URL change that becomes a habit

QwenCloud “supports industry-standard protocols, including chat completions and responses APIs compatible with OpenAI’s specification, as well as an API interface compatible with Anthropic.” In practice: anyone with an existing OpenAI integration, or with Claude Code, changes one line — the base URL — and has Qwen3.8-Max in production by the afternoon. The page already shows the ready-made configuration for Claude Code, with ANTHROPIC_BASE_URL pointed at an Alibaba domain.

That’s also the point of risk. A base-URL change is reversible in theory and sticky in practice. A technical report states whatever it wants to state, and stays silent on the rest: here the post declares benchmarks and use cases across dozens of sections, and says nothing at all about where data goes during hosted inference, how long it’s kept, who can access it. Anyone who starts sending documents to QwenCloud while waiting for the weights is making a decision about their data, not about the model — and making it without really deciding it, one prompt at a time, until it becomes the team’s habit.

The other side, for fairness

Opening the weights of a Max-class model is a significant fact: if it genuinely happens, it’s good news for anyone seeking independence from cloud vendors — a model that can run in-house, without depending on an external API, is the direction we recommend. This piece’s criticism isn’t aimed at Qwen or at China: it’s aimed at the way an announcement — “the weights will open next week” — gets read as though it were already available. The same applies to any Western lab that promises weights “coming soon”: the right question isn’t who says it, it’s whether the file exists.

What to do about it

  • Don’t route confidential documents through the hosted API “while waiting for the weights”: that’s an architectural decision, and it should be made as one.
  • Wait for the licence text before planning adoption, not for the weights announcement: they are two separate events.
  • When the text appears, read it like a supply contract: revenue or user thresholds, field-of-use restrictions, revocation conditions, liability for third-party rights.
  • Verify what is actually published — weights, training recipe, data, checkpoints — before writing “open-weight model” into a specification: today, for Qwen3.8-Max, that phrase is not yet verifiable.

How we solve this

Verifying the licence, provenance and openness scope of every model before it goes into production isn’t advice that ends up filed in a document: it becomes a check that runs on the client’s own systems, with a register of installed models — licence, version, verification date, origin of the weights — ready to show an inspector and updated with every new release.

The same setup brings the organisation’s scattered data — archives, business systems, documents, sensors, plant — into a single operating model on which AI agents execute decisions with a human operator in command: for large enterprises, defence, government and healthcare. Compliance on the model is the entry point; the decision-making system running on top of it is what we sell. In both delivery modes — on-premise on the client’s self-contained machines, or a dedicated cloud with a dedicated VPN and a data centre in Italy, in premises we staff directly — and always with shared governance: many clients don’t have anyone in-house to administer AI models, and they don’t need to start now.

Assessing Qwen3.8-Max, or another model with weights “on the way”, before the licence text is public? Thirty minutes with one of our experts: we’ll put in writing what’s missing today and what to check once the file genuinely appears.

Sources