Review Machine · Master Review · AI models
Qwen3.8-2.4T-A95B
The charts are the launch, the download is the checkpoint
The open checkpoint behind Qwen3.8-Max is documented as text-only with thinking always on, frozen at its August state while the hosted model took a September post-training pass, and no independent evaluation of the weights themselves was readable on 20 September 2026.
Open weights published 12 August 2026, Hugging FaceArchitecture 2.4T total / 95B active, 92 layers, 512 expertsHosted price $2 in / $6 out per 1M tokens, QwenCloudLicence Qwen3.8-Max LicenseRead 20 September 2026Skip to the verdictRead as text
Verdict age
ProvisionalVendor pages settled · independent runs thinThe weights are five weeks old as we read, and the hosted model took a September post-training pass that has no open checkpoint, so a later checkpoint or a first independent run would change this reading.
- Early · editors only
- Provisional · editors in, owners thin
- Settled · owner reviews read over months
From 19 sources to one verdict
Seven levels, in the same order in every Master Review. Each line is the takeaway; open a level for the evidence under it, or stop when you have enough.
What we read
19 sources, 3 Aug – 20 Sep 2026. Forum and buyer tiers thin. Editorial tier empty.
By tier, heaviest first
The model card, the licence file, the launch post, the QwenCloud page and the repository, read 20 September 2026.
One page, and it scores a September build of the hosted model rather than these weights.
Counts from one unauthenticated GitHub REST API request, not ratings.
Four threads from one Hacker News API search that answered 94 stories.
Eight write-ups about the release, read 20 September 2026; several sit on vendor-adjacent blogs, so their claims are reported as theirs.
Date window
3 Aug 2026 to 20 Sep 2026
Editorial reviews 20 Sep 2026Official pages read 20 September 2026; the launch post is dated 3 August 2026.
Couldn’t read, so not used
- South China Morning Post report on the releasepaywall
- Qwen3.8-Max Overview page on the forum announcementno dated page of its own could be opened
Nothing that rests only on a source we could not read is counted, and the paywalled report is not cited.
Read, then rejected
- Thomas Wiegold's hands-on review · a different version, the July preview, and it reports running the model
- Trilogy AI's StackPerf run · a different version, one matched run
- DataCamp's benchmark breakdown · restates the maker's table for the hosted build
What they measured
2.4 trillion parameters with 95 billion active per token, a native 262,144-token context, and 4,142 stars on the repository five weeks after the weights landed.
Parameters
2.4trilliontotal, with 95 billion active per token, 92 layers and 512 experts
Hugging Face · read 20 SepNative context
262,144tokensthe card says it is extensible to 1,010,000
Hugging Face · read 20 SepWeight file size
4.9TB BF16the FP8 build is put near 2.4 TBA provider's own write-up, not the card
OrcaRouter · read 20 SepRepository figures
4,142starswith 314 forks and 20 open issues
GitHub REST API · read 20 SepTerminal-Bench 2.1
Vendor table. The card says its own entries ran on a coding harness while the rivals' cells are their best published scores.
Output speed
38.6tokens/sthe independent page calls this slow; it is the September hosted build, not the weightsDifferent version
Artificial Analysis · read 20 SepWhere they agree
The checkpoint is the Max-class core, documented consistently, with no independent run of the weights themselves.
One tick per editorial review, in order of publication:AA Artificial Analysis
The downloadable checkpoint is the Max-class core
0/1The card, the launch post and seven write-ups give the same architecture and the same limits. T1Official Other
The licence grant is broad below two revenue lines
0/1Use, modification, hosting and fine-tuning are granted free of charge; the conditions begin at 100 million monthly users or US$20 million monthly revenue, and at US$50 million for Model-as-a-Service and AI-work-assistant businesses. T1Official Other
The front group, on the maker's own table
0/1Terminal-Bench 2.1 at 86.6 and PaperBench at 93.0, both vendor figures whose footnotes say Qwen's own entries ran on a coding harness while the rivals' cells are published scores from elsewhere, and one audit that reads the model as belonging in the frontier cohort. T1Official Other
A fixed artifact you can pin
0/1Two write-ups put custody and reproducibility first as the reason to take the weights over a rolling endpoint. Other
Attention, which says little on its own
0/14,142 stars, 314 forks and 20 open issues on the repository on 20 September 2026, and one API search that answered 94 stories about the family. T3Buyers Forum
What keeps coming up against it
The version you can download is not the version being measured, and the licence is the part with the longest tail.
The version you can download is not the version being measured
Well documented · count not publishedThe card describes the checkpoint as text-only with thinking required for all interactions, and points to the hosted version for vision, a non-thinking mode, a 1M window and built-in tools.
- Hugging Face a text-only model that requires thinking mode for all interactions
Hugging Face · read 20 Sep
The hosted page lists image and video input, function calling, structured outputs and a 1M context, none of which the checkpoint's card claims.
QwenCloud · read 20 Sep
Three write-ups state the same split, one of them describing the hosted product as the official version built on these weights with more features.
OrcaRouter · read 20 SepAI Tools Recap · read 20 SepGroundy · read 20 Sep
The weights carry a separate licence with revenue conditions
0/1The shipped file is titled the Qwen3.8-Max License, not the Apache 2.0 of the previous generations, and six write-ups report the same. T1Official Other
Serving the checkpoint is a rack-scale job
0/1BF16 weights put near 4.9 TB, an FP8 build near 2.4 TB, and a serving floor the write-ups place at multi-node accelerators. Other
Thinking cannot be switched off in the download
0/1The card states that thinking cannot be disabled, and the hosted version offers a non-thinking mode. T1Official Other
Where they split
Five places the evidence pulls two ways, starting with the date the weights landed.
The date the weights landed
The thread announcing the checkpoint is dated 12 August 2026.
Against that: Dates the open release 13 August 2026.
Our readWe take 12 August, the date on the checkpoint's own thread.
The licence the repository shows against the licence the weights carry
The repository is listed under Apache 2.0 on 20 September 2026.
Against that: The licence shipped with the weights is the Qwen3.8-Max License, with two revenue conditions.
Our readBoth accurate about different things: the repository's file covers the series, the shipped licence governs the weights.
Whether the open checkpoint matches the hosted model
Reports the independent index scoring the hosted build and the open checkpoint at the same aggregate figure.
Against that: States that no independent benchmark of the downloadable checkpoint has been published, and that the hosted build carries a September post-training the weights do not.
Our readA matching aggregate on one index does not make two artifacts one product, and the version gap is most of the difference.
What the launch table measures against
Its own entries ran on a coding harness, while the rivals' cells are their best published scores from elsewhere.
Against that: Scores the September hosted build at 45 on its intelligence index, its own number for a different version.
Our readNeither is a head-to-head, and no evaluation we could read scores the checkpoint itself.
Where the developer threads went
The launch post drew 1,124 points and 612 comments, the checkpoint 713 points and 171.
Against that: Among the stories the one search returned, the release thread after the checkpoint is the Flash-Next one, dated 26 August 2026, and the September threads are about the smaller sibling's quantisations.
Our readThe flagship's day was the launch, and the weeks after belong to the models people can run.
Who it's for, who should pass
For a team that needs the weights in its own custody and reads the release for what it is. Pass if you wanted the model the hosted version is.
It suits you if
- You need the weights in your own custody, on hardware you already operate.The licence covers hosting and fine-tuning below its revenue lines, and a fixed checkpoint can be audited and pinned.
- Your work is long-horizon agentic coding and text input is enough.The card and the launch post are written around coding and long tasks, and the one performance figure we could read puts the hosted build at 38.6 output tokens per second and calls that slow.
- You are reading the release rather than shopping for a product.A documented checkpoint with a published licence is worth studying on its own terms, even where the product around it is not.
Pass if
- You wanted the model the hosted version is.The card describes the weights as text-only with thinking always on, while the hosted product adds image and video input, a non-thinking mode, a 1M window and built-in tools.
- You have no multi-node GPU capacity.Three write-ups put the serving floor at rack scale and recommend the smaller sibling for one machine.
- You need unconditional open weights.The previous generations shipped under Apache 2.0; these weights carry the Qwen3.8-Max License, and the file itself is the operative text.
This is a reading of published reviews, not medical, financial or legal advice. For a decision about your health, your money or your rights, a qualified professional is the right next step, and not a review.
The verdict
Vendor documentation settled, no independent run of the weights, and a version gap that is the whole story.
A genuine Max-class checkpoint with a documented architecture and a permissive grant, published under a licence with revenue conditions, and frozen at its August state while the hosted model moved on, with no independent evaluation of the weights themselves readable.
Confidence, by tier
The card, the licence file, the launch post, the model page and the repository, all read on 20 September 2026.
The one evaluation page we could open scores a September build of the hosted model, not these weights.
Four threads from one API search that answered 94 stories, weighted to the launch week.
One GitHub API call for stars, forks and open issues, which measure attention and not quality.
- Editorial evidence
- under a month old
- Newest report
- None read
- Read
- 20 Sep · month 1
You can download the max-class weights. The max-class model stays on the API.
Rests onThe model card describes the checkpoint as text-only with thinking that cannot be disabled, against the hosted version's vision, non-thinking mode and 1M window, and a provider's write-up of 20 September 2026 reports that the September post-training has no open checkpoint.
Sources
19 sources, heaviest tier first. Every figure above comes from one of these.
- T1OfficialQwen3.8-Max: A New Bar for Coding and Cowork3 Aug
- T1OfficialQwen3.8-2.4T-A95B model cardread 20 Sep
- T1OfficialQwen3.8-Max License, shipped with the weightsread 20 Sep
- T1OfficialQwen3.8-Max model pageread 20 Sep
- T1OfficialQwenLM/Qwen3.8 repository and READMEread 20 Sep
- T2EditorialQwen3.8 Max (0902)read 20 Sep
- T3BuyersRepository figures for QwenLM/Qwen3.8one unauthenticated request for stars, forks and open issuesread 20 Sep
- ForumQwen3.8-Max: A New Bar for Coding and Cowork, 1,124 points, 612 comments3 Aug
- ForumQwen3.8-2.4T, 713 points, 171 comments12 Aug
- ForumQwen3.8-Flash-Next, 704 points, 233 commentsa different release of the same family, cited only in Where reviewers split26 Aug
- ForumSearch for Qwen3.8, 94 storiesone search request that reports a total, no pagingread 20 Sep
- OtherQwen3.8 Max release auditread 20 Sep
- OtherQwen3.8-Max against the open weightspublished by a provider that resells model accessread 20 Sep
- OtherOpen weights under a custom licenceread 20 Sep
- OtherTwo licences for one releaseread 20 Sep
- OtherThe first Max-class open weightsread 20 Sep
- OtherThe custom licence decodedread 20 Sep
- OtherWeights against the APIread 20 Sep
- OtherQwen 3.8 review and hardwareread 20 Sep
MethodOn 20 September 2026 we read 19 sources dated 3 August to 18 September 2026, from the model card, the licence file and the launch post to one independent evaluation page, eight write-ups and four Hacker News threads, and synthesised them with AI. Nothing was downloaded, run or tested.
Dates without a year are 2026.
The review as text
Qwen3.8-2.4T-A95B is the open-weight checkpoint behind Qwen3.8-Max, Alibaba's 2.4-trillion-parameter mixture-of-experts flagship, and the first Qwen-Max-class model the company has published as downloadable weights. The launch post went up on 3 August 2026 and promised the weights "next week". The checkpoint reached Hugging Face on 12 August 2026, while the hosted Max-class product stayed on sale through QwenCloud with features the download does not carry.
This reading covers nineteen sources read on 20 September 2026: five pages published by Alibaba or Hugging Face, one independent evaluation page, eight write-ups, four Hacker News threads from one API search that answered 94 stories, and one GitHub API call for the repository's figures. The material dates from 3 August to 18 September 2026. No independent evaluation of these weights was readable, and the news report we found was paywalled, so both gaps are named below rather than papered over.
Consensus
What the sources agree on is what the checkpoint is, not how well it works. Its shape and limits are documented consistently: 2.4 trillion parameters in total with 95 billion active per token, 92 layers, 512 experts, a native context of 262,144 tokens that the card says can be extended to 1,010,000, and text as its only input. Nine of the nineteen sources, two official pages and seven write-ups, restate that description. On capability the agreement stops. The only independent evaluation we could open scores a September build of the hosted model, not these weights, and no source we read claims to have run the checkpoint itself. Confidence here is thin, and the reason is the version boundary rather than a disagreement between sources.
Recurring strengths
The grant is broad below its two revenue lines. The licence file shipped with the weights grants use, modification, publication, distribution, sale, hosting, fine-tuning and derivative works free of charge. Seven of the nineteen sources, that licence file and six write-ups, describe it that way, and the conditions bite only at scale. Attribution is required above 100 million monthly active users or US$20 million monthly revenue, and a separate licence above US$50 million for Model-as-a-Service and AI-work-assistant businesses.
The download is an artifact you can pin. Two of the write-ups put custody and reproducibility first, the practical case for a versioned checkpoint over a rolling endpoint.
The maker's table puts the family in the front group. The card's Terminal-Bench 2.1 row reads 86.6 for Qwen3.8-Max against 88.8 for GPT-5.6 Sol and 84.6 for Claude Opus 4.8, and its PaperBench row, 93.0, is the highest in that table. Three of the nineteen sources, the card, the launch post and one audit, place the Max-class build there. Its footnotes say its own entries ran on a coding harness while the rivals' cells are published scores from elsewhere.
Attention is easy to measure and says little. The repository held 4,142 stars with 314 forks and 20 open issues on 20 September 2026, five weeks after the weights landed, and one API search answered 94 stories about the family, weighted to the launch week.
Recurring complaints
A separate licence, not the Apache 2.0 of the previous generations. Seven of the nineteen sources, the licence file and six write-ups, report that the weights carry a document titled the Qwen3.8-Max License, and the write-ups read its second clause as aimed at providers who resell access. Whether that licence permits a reader's own use is a legal question this piece does not answer.
The serving floor is a rack. Three of the write-ups put the BF16 weights near 4.9 TB, an FP8 build near 2.4 TB, and the smallest documented serving configuration at multi-node accelerators, which is why the same write-ups recommend the smaller sibling instead.
Thinking cannot be switched off in the download. Four of the nineteen sources, the model card and three write-ups, report that every response begins with reasoning and that the mode cannot be disabled, while the hosted version offers a non-thinking mode.
This is a reading of published reviews, not medical, financial or legal advice. For a decision about your health, your money or your rights, a qualified professional is the right next step, and not a review.
Where reviewers split
The date the weights landed. The thread announcing the checkpoint is dated 12 August 2026, and one write-up dates the release 13 August. We take 12 August, the checkpoint's own thread date.
The licence the repository shows against the licence the weights carry. The GitHub repository is listed under Apache 2.0, and the file shipped with the weights is the Qwen3.8-Max License. Both are official pages read the same day and accurate about different things. Our read: the repository's file covers the series, the shipped licence governs the weights.
Whether the open checkpoint matches the hosted model. One audit reports that the independent index scored the hosted build and the open checkpoint at the same aggregate figure. A provider's write-up states that no independent benchmark of the downloadable checkpoint has been published, and that the hosted build carries a September post-training the weights do not. Both true: a matching aggregate on one index does not make two artifacts one product, and the version gap is most of the difference.
What the launch table measures against. The card's footnotes say its entries ran on one harness and the rivals' cells elsewhere. The independent page we could open scores the September hosted build at 45 on its intelligence index, its own number for a different version. Neither is a head-to-head, and no evaluation we could read scores the checkpoint itself.
Where the developer threads went. The launch-post thread drew 1,124 points and 612 comments, and the checkpoint's own thread 713 points and 171. Among the twenty stories that search returned, the later ones are about the smaller sibling's quantisations and the Flash-Next release. The flagship's day was the launch, and the weeks after belong to the models people can actually run.
Who it suits
You need the weights in your own custody, on hardware you already operate. The licence covers hosting and fine-tuning below its revenue lines, and a fixed checkpoint is the thing that can be audited and pinned.
Your work is long-horizon agentic coding and text input is enough. The card and the launch post are both written around coding and long tasks, and the one performance figure we could read is a warning about patience: the independent page puts the hosted build at 38.6 output tokens per second and calls that slow.
You are reading the release rather than shopping for a product. A documented checkpoint with a published licence is worth studying on its own terms.
Who should pass
You wanted the model the hosted version is. The card describes the weights as text-only with thinking always on. The hosted product adds image and video input, a non-thinking mode, a 1M window and built-in tools, and five of the eight write-ups say the two are not the same artifact.
You have no multi-node GPU capacity. Three of the write-ups put the serving floor at rack scale and recommend the smaller sibling for one machine.
You need unconditional open weights. Previous Qwen generations shipped under Apache 2.0. These weights do not, and the file itself is the operative text. One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it. Get the weekly verdict.
Sources
- Qwen3.8-2.4T-A95B model card, official, read 20 September 2026.
- Qwen3.8-Max License, shipped with the weights, official, read 20 September 2026.
- Qwen3.8-Max: A New Bar for Coding and Cowork, Qwen blog, official, published 3 August 2026.
- Qwen3.8-Max, QwenCloud model page, official, read 20 September 2026.
- QwenLM/Qwen3.8 repository, official, read 20 September 2026.
- Repository figures for QwenLM/Qwen3.8, the public GitHub REST API, one request, 20 September 2026.
- Qwen3.8 Max (0902), Artificial Analysis, independent evaluation, read 20 September 2026.
- Hacker News API search for Qwen3.8, 94 stories, read 20 September 2026.
- Hacker News, the launch post, forum, 3 August 2026.
- Hacker News, the checkpoint, forum, 12 August 2026.
- Hacker News, the Flash-Next release, forum, 26 August 2026.
- Groundy, Qwen3.8 Max release audit, other, read 20 September 2026.
- OrcaRouter, Qwen3.8-Max against the open weights, other, read 20 September 2026.
- Machine Brief, open weights under a custom licence, other, read 20 September 2026.
- SQ Magazine, two licences for one release, other, read 20 September 2026.
- The Ledger, the first Max-class open weights, other, read 20 September 2026.
- Latent East, the custom licence decoded, other, read 20 September 2026.
- AI Tools Recap, weights against the API, other, read 20 September 2026.
- AirMore, Qwen 3.8 review and hardware, other, read 20 September 2026.
On 20 September 2026 we read 19 sources dated 3 August to 18 September 2026, from the model card, the licence file and the launch post to one independent evaluation page, eight write-ups and four Hacker News threads, and synthesised them with AI. Nothing was downloaded, run or tested.
One verdict a week.
Every week, the most useful Master Review we finished: what the internet agrees on, the complaint that kept appearing, and who should skip it.