What actually shipped
When we covered the Qwen3.8 Max preview in July, the single most consequential unverified line was the open-weight promise. As of August 2026, that promise has been honoured — but as two releases with two very different sets of terms, and the difference between them will decide which one your business can touch.
On 12 August 2026, Qwen3.8-2.4T-A95B — the open checkpoint of the Max flagship — landed on Hugging Face and ModelScope: 2.4 trillion total parameters with roughly 95 billion active per token in a fine-grained mixture-of-experts design. It is, as far as we can establish, the first time a Qwen-Max-class flagship has shipped weights at all, and it arrived under a new custom licence — named simply qwen3.8-max in the repositories — rather than the Apache 2.0 that previous Qwen generations used.
On 14 August, the companion Qwen3.8-27B followed: a dense 27-billion-parameter model, released under Apache 2.0, with native vision-language input included. Where the flagship needs a cluster, the 27B fits a single high-end GPU — early guides put it at around 17GB at Q4_K_M quantisation, with 4-bit VRAM estimates in the 14–16GB range.
Two models, two days apart, two licences. One is open-weight. The other is open-source. That distinction is not pedantry — it is the difference between a model your legal team waves through and one it schedules a meeting about.
The flagship: what you get, and what got stripped
Start with what the 2.4T release actually contains, because it is not the hosted Qwen3.8-Max in a downloadable box. The open checkpoint is text-only — the vision input available on the hosted service was not included. Its native context window is 262,144 tokens, extensible towards roughly 1M with long-context techniques, whereas the hosted version defaults to 1M tokens outright. Early deployment guides report the BF16 weights at roughly 4.9TB, shipped across more than 200 safetensors shards, with an official FP8 variant alongside; even the most aggressive community quantisations still occupy hundreds of gigabytes. This is multi-node infrastructure, full stop — the same class of problem we mapped in our guide to serving trillion-parameter MoE models on your own hardware.
The gap between the label and the artefact has not gone unnoticed. Per explainx.ai's release coverage, a Hugging Face discussion opened within hours calling the stripped feature set a disappointment relative to what the Max branding implies. That reaction is worth tempering: a downloadable near-frontier 2.4T checkpoint is still a first for a Qwen flagship, and the hosted service — which reached general availability in early August, per multiple release guides — remains the way most teams will actually consume the flagship. But if you were expecting the July preview's full capability sheet in a tarball, you did not get it.
The licence, clause by clause
Here is where the release stops being a model story and becomes a business story. The qwen3.8-max licence is not Apache 2.0, is not OSI-approved, and therefore — whatever the launch threads say — the flagship is not open-source. Per the terms reported consistently across multiple licence analyses, it works like this:
- Attribution threshold: products with more than 100 million monthly active users, or more than $20M in monthly revenue, must display the model name prominently in their user interface.
- Separate-licence threshold: model-as-a-service or "AI work assistant" businesses exceeding $50M in aggregate revenue in any 12-month period must obtain a separate licence from Qwen before using the software.
- Internal-use carve-out: deployments not exposed to third parties are exempt from the commercial conditions entirely.
What that separate licence costs is the open question. Coverage from explainx.ai reports it as a revenue-sharing arrangement whose exact percentage was still being finalised at release — which, if accurate, would make this the first major open-weight release with a revenue-share mechanism built into its commercial tier. We could not confirm the percentage against the licence text itself, so treat that characterisation as reported rather than settled. Either way, the structural point stands: above a defined revenue line, your right to serve this model commercially depends on a negotiation with Alibaba.
The reported licence text does not clarify whether the $50M threshold applies to your total company revenue or only to revenue from the AI service built on the model. If you are anywhere near that line — or plan to be — get counsel to read the actual LICENSE file in the repository before anything built on the 2.4T checkpoint touches production. Precedent from Qwen's earlier Apache releases tells you nothing about this document.
For most readers of this site, the honest summary is that the thresholds are generous: a startup in Pune or Leeds serving the model to customers is nowhere near 100M MAU or $50M in annual AI revenue. But licences are read at acquisition time, at enterprise procurement time, and at investor due-diligence time — and "custom Chinese-vendor licence with an unpriced revenue-share tier" is a very different diligence conversation from "Apache 2.0".
The 27B is the release most builders will actually use
Which brings us to the quieter release of the pair, and arguably the more important one. Qwen3.8-27B is dense rather than sparse, ships under unmodified Apache 2.0, and — unlike its 2.4T sibling — kept its vision-language input. One analysis lists text, image and video input with a 262K native context extendable via YaRN; at around 17GB in Q4_K_M form, per early guides, it runs on a single 24GB GPU or an Apple Silicon workstation. It is the natural heir to Qwen 3.6's 27B, which put serious coding capability on a consumer GPU — except this generation adds vision and arrives into a far more mature local-inference ecosystem.
Alibaba's own published benchmarks put the 27B at 61.7% on SWE-bench Pro against the flagship's 67.7% — vendor numbers, unverified independently, but the shape of the claim is the point: the licence-clean, workstation-sized model is claimed to land within striking distance of the 2.4T cluster model on the workloads most teams care about. If that survives independent evaluation, the value-for-constraint trade is not close.
| Qwen3.8-2.4T-A95B | Qwen3.8-27B | |
|---|---|---|
| Released | 12 August 2026 | 14 August 2026 |
| Architecture | 2.4T total / ~95B active MoE | 27B dense |
| Licence | Custom qwen3.8-max (not open-source) | Apache 2.0 (open-source) |
| Modality | Text-only | Vision-language included |
| Native context | 262K (extensible towards ~1M) | 262K, per early analyses |
| Hardware class | Multi-node cluster (~4.9TB BF16, per guides) | Single 24GB GPU (~17GB at Q4_K_M, per guides) |
| Commercial conditions | MAU/revenue attribution; separate licence above $50M | None |
Which one can you build a business on?
If you buy hosted API capacity
Nothing about the licence split changes your position except leverage — and it improves it. A downloadable near-frontier checkpoint, even one almost nobody will self-host, resets the ceiling on what closed API vendors can charge, exactly as DeepSeek V4-Pro's break-even maths did earlier this year. Run the hosted Qwen3.8-Max against your evaluation set, file the results, and bring them to your next pricing conversation — whether that is an Indian cloud commitment renewal or a UK enterprise procurement round.
If you self-host
The 2.4T checkpoint is, for almost everyone reading this, a strategic artefact rather than a deployment target. The teams that can fund a multi-node cluster can also fund the legal review the licence demands; if that is you, the question is whether a text-only 262K-context flagship beats the open MoE alternatives you can already serve. For everyone else, the 27B is the release: Apache 2.0 means no user or revenue thresholds and no separate-licence cliff, only the standard notice-retention obligations — and a footprint that makes a production vLLM deployment a weekend project rather than a capital expenditure. If you are still weighing the build-versus-buy line itself, our self-host or API decision guide covers the maths that stays constant across model generations.
The dual-market compliance angle cuts the same way in both countries. For UK teams, a self-hosted Apache 2.0 model inside your own infrastructure is usually the easiest conversation you will ever have with a data-governance review — no cross-border processing, no vendor terms to assess. For Indian teams handling personal data under the DPDP regime, the same logic applies. The custom-licensed flagship, by contrast, adds a commercial-terms review on top of the usual security one, and the unpriced separate-licence tier is precisely the kind of open-ended obligation enterprise counsel flags.
Before any open-weight model enters your stack, record three things in your engineering log: the exact licence file hash at the commit you downloaded, the date, and the thresholds that apply to you. Licences on hosted repositories can change; your rights generally attach to the version you took delivery of. Five minutes of provenance now saves a very expensive argument later — and a documented model-governance habit is exactly the kind of artefact worth showing on a Builder profile.
What this does to the rest of the open-weight market
The split-licence structure is the part every other lab will study. Alibaba has effectively priced the frontier tier while giving away the practical tier — capturing goodwill and ecosystem share with the Apache 27B, while keeping a commercial handle on anyone who builds a large business on the flagship. It is a more surgical version of the threshold clauses Meta pioneered with the Llama line, and it lands in a market where Moonshot's 2.8T Kimi K3 and Z.AI's GLM-5.2 have kept genuinely permissive terms — for now. If the revenue-share tier proves enforceable and lucrative, expect the next round of leaderboard flips to come with more lawyers attached.
There is also a definitional battle underway, and it matters that builders fight their corner of it. Every headline that calls the 2.4T release "open-source" erodes the term a little further. The 27B is open-source. The flagship is open-weight under conditional commercial terms. Using the words precisely is not pedantry — it is how you keep the next licence honest.
The bottom line
As of August 2026, the answer to "which of these can I build a business on?" is unambiguous for most teams in India and the UK: the 27B, today, with no caveats a startup needs to care about. The 2.4T flagship is a landmark — the largest checkpoint Qwen has ever made downloadable — but it is a landmark you visit, not one you move into: text-only, cluster-scale, and wrapped in a licence whose commercial tier is still being priced. The preview promised open weights; the release delivered them, and in doing so demonstrated that "open" is no longer one thing. Read the LICENSE file first. Then read it again with the person who signs your contracts.
Primary sources: the Qwen model repositories on Hugging Face carry the weights and the licence text for both releases — read the LICENSE file in each repository rather than any summary of it, including this one. Release specifications and licence thresholds above are drawn from explainx.ai, MindStudio, SQ Magazine and Codersera; the benchmark figures are Alibaba's own published numbers and have not been independently verified.