Qwen 3.8: Alibaba’s 2.4T Model Tested Against the Fable 5 Claim

Qwen 3.8 is no longer a preview with a marketing claim attached to it. Alibaba made Qwen3.8-Max generally available on August 3, 2026, and this time the model arrived with the numbers it was missing in July: 95 billion active parameters per token out of 2.4 trillion total, a confirmed 1-million-token context window, and API pricing at $2 per million input tokens and $6 per million output tokens.

That matters because the claim Alibaba led with, that Qwen 3.8 is “second only to Fable 5,” is finally testable. It turns out to be true on one leaderboard and wrong on another, and the gap between those two answers tells you more about the model than the headline did. This article covers the confirmed specs, the real pricing, where the open weights actually stand, and how the Fable 5 claim holds up against independent rankings.

The Key Takeaways

• Qwen3.8-Max went generally available on August 3, 2026, about two weeks after its July 19 preview.
• It runs 2.4 trillion total parameters with only 95 billion active per token, so its real compute cost sits closer to a mid-sized model than the headline number suggests.
• API pricing is $2 input and $6 output per million tokens, with cached reads at $0.25.
• The “second only to Fable 5” claim is correct on Arena’s vision board (#2) and wrong on its text board (#5).
• Open weights shipped in mid-August 2026: the text-only Qwen3.8-2.4T-A95B under a custom licence, and a smaller Qwen3.8-27B under Apache 2.0.

What Is Qwen 3.8?

Qwen 3.8 is the current flagship of Alibaba Cloud’s Qwen family, shipping as Qwen3.8-Max. It is a sparse Mixture-of-Experts (MoE) model with 2.4 trillion total parameters, of which roughly 95 billion activate on any given token. That second figure was the biggest gap in the July preview, and it reframes the model entirely; the per-request compute burden is closer to a mid-sized dense system than to anything you would expect from a 2.4-trillion-parameter headline.

The other headline is what Qwen 3.8 can see. It accepts text, images, and video as input and returns text, making it the first Qwen model above 1 trillion parameters to go multimodal. The context window is now confirmed at 1 million tokens, with a practical ceiling of 991,000 input tokens (983,000 with thinking enabled) and 131,000 output tokens.

Compare that table to the one Alibaba shipped in July and the difference is stark. Every row that read “not disclosed” now has a number behind it, which is the main reason the model is worth a second look. The launch details were laid out in MarkTechPost’s release coverage.

Qwen 3.8 Pricing and How to Access It

The preview-era answer to “what does it cost” was a shrug and a discount. That is over. Qwen3.8-Max now has a published per-token rate, and it undercuts most Western frontier models by a wide margin while sitting slightly above its own predecessor.

The caching tiers are worth reading carefully if you plan long-context workloads. At $0.25 per million for implicit cached reads, re-sending a large document across many turns costs a fraction of the first pass, which is the whole point of pairing a 1-million-token window with aggressive cache pricing.

Where you can use Qwen3.8-Max

Access changed shape at launch. The model is live for global developers through Alibaba Cloud’s Model Studio APIs, and for everyone else through QwenWork, Alibaba’s workplace AI agent platform, which entered public beta on August 2, 2026.

If you tried the July preview through QoderWork, note that it no longer exists as a standalone product. Alibaba folded QoderWork, MuleRun, and Wukong into the single QwenWork platform, so that is where those workflows now live. SCMP’s launch report covers the rollout across both surfaces.

One caveat for business users. QwenWork routes enterprise workflow data through a Chinese provider, which brings the usual state-law and data-residency questions with it. That is a procurement conversation, not a technical one, but it is worth having before you wire it into anything sensitive. If you are weighing frontier models against each other first, our guide on which AI model to use for each task lays out the trade-offs.

Is Qwen 3.8 Really “Second Only to Fable 5”?

Here is the answer the July article could not give you. Qwen3.8-Max is now ranked on Arena, the blind human-preference leaderboard, and the claim splits cleanly in two. On the vision board it is correct. On the text board it is not.

On vision, the claim holds

Qwen3.8-Max sits at #2 on Arena’s vision leaderboard with a score of 1305, behind only claude-fable-5 at 1318. That is exactly the ranking Alibaba described, and it makes Qwen 3.8 the highest-ranked non-Anthropic model for multimodal work. For a first multimodal flagship, that is a genuinely strong debut.

On text, it lands fifth

The text board tells a different story. Qwen3.8-Max ranks #5 with 1496, behind four Anthropic entries rather than one.

Read that table carefully before you dismiss the result. Qwen 3.8 is one point behind claude-opus-4-6 and thirteen behind the leader, which is close enough that the ordering could shift. It is also the top-ranked Chinese model on the text board, which is a real milestone even if it is not the one Alibaba advertised.

Two honest caveats. Both Arena scores are still flagged preliminary, carrying wider error bars than established entries, so they will move as votes accumulate. And Arena measures blind human preference, not correctness, so it rewards models that produce answers people like reading. Treat it as one strong signal rather than a verdict.

Qwen 3.8 Benchmark Results

Alibaba published a full benchmark table at general availability, which is the single biggest change from the July preview. The pattern in the numbers is more interesting than any individual score; Qwen 3.8 is exceptional at research and document work and merely competitive at software engineering.

The soft spot is at the bottom of that table. On DeepSWE 1.1 Qwen 3.8 scores 56.6 against Fable 5’s 70.0, a gap of more than thirteen points on agentic software engineering. If your use case is autonomous coding, that single row matters more than the five scores above 85.

Alibaba also leaned on two long-horizon demonstrations rather than benchmarks alone. In one, the model spent over ten days autonomously building a self-evolving software harness, running its own tests and iterating without human input. In another it reproduced a machine learning paper from scratch, running 33 rounds of GPU training across roughly 125 hours and writing about 7,600 lines of code. Those are vendor-run demos, not independent evaluations, so weigh them accordingly.

One number to keep separate. Qwen3.8-Max has not been scored on the Artificial Analysis Intelligence Index yet, so anyone quoting it an index number is guessing.

Where the Open Weights Actually Stand

This is the part developers care about most, and it is the one promise that took an extra week to settle. Alibaba said the weights would ship the week of August 10, 2026, alongside a second, much smaller Qwen3.8-27B checkpoint aimed at people who cannot run a 2.4-trillion-parameter model.

Both arrived. The flagship weights went up on the official Qwen organisation on Hugging Face as Qwen3.8-2.4T-A95B around August 13, 2026, with an FP8 copy alongside, and Qwen3.8-27B followed within days. The download is not the full API model. Alibaba’s own model card says Qwen3.8-Max adds vision input, a non-thinking mode and a 1M context by default on top of it, while the open checkpoint is text-only with 262,144 tokens natively. It is still the first Max-class Qwen ever released as open weights.

One practical warning. A third-party repository named Qwen3.8_4B_Distilled collected thousands of downloads on Hugging Face before the real release, but its own metadata identifies it as a fine-tune of Qwen3-4B-Thinking, an older and much smaller model. It is not an official Qwen 3.8 release and it is not a distillation of one. Download only from the official Qwen organisation.

The licence is where the two checkpoints split. Qwen3.8-27B ships under plain Apache 2.0. The flagship uses a custom Qwen3.8-Max licence: free to use, modify and sell, but products above 100 million monthly users or $20 million in monthly revenue must display the model name, and model-as-a-service or coding and office assistant businesses above $50 million a year need a separate licence. For almost every self-hoster, neither condition applies. Our guide to the best open-source AI models covers what else you can download today. Tencent cleared the same bar, releasing Hy4 Preview under Apache 2.0 on August 28, 2026.

Qwen 3.8 vs the Competition

Qwen 3.8 landed in the middle of a rush by Chinese labs into the multi-trillion-parameter tier. Now that it has real pricing and real scores, the comparisons are finally concrete instead of speculative.

Qwen 3.8 vs Kimi K3

Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on July 16, 2026, three days before Qwen 3.8’s preview. K3 is larger on paper, and it shipped its weights almost four weeks before Qwen 3.8’s arrived.

Pricing now separates them clearly. Qwen 3.8 costs $2 / $6 against Kimi K3’s $3 / $15, making Qwen roughly two and a half times cheaper on output. Qwen 3.8 also handles video input, while K3 is limited to text and images. For the deeper breakdown of Moonshot’s model, see our coverage of Kimi K3’s open-weight launch.

Qwen 3.8 vs Qwen 3.7-Max

The generational jump is now measurable. Qwen3.7-Max scored 56.6 on the Artificial Analysis Intelligence Index in May 2026, a fifth-place result at the time, and it was text-only. Qwen 3.8 adds image and video input, a confirmed 1-million-token window, and a top-five Arena placement.

Whether the upgrade costs you more depends entirely on which Qwen3.7-Max price you compare against. Qwen 3.8 runs $2 / $6. Qwen3.7-Max is on a 50% promotional discount that puts it at $1.25 / $3.75 against a $2.50 / $7.50 list price. Measured against that promo, Qwen 3.8 output is about 60% more expensive; measured against list, it is about 20% cheaper. Alibaba has published no end date for the discount, so for text-only workloads where the older model already performs well, check which of those two numbers you will actually be billed before you migrate.

Qwen 3.8 vs Claude Fable 5

This is the comparison Alibaba invited, and the split verdict above is the honest answer. On multimodal work Qwen 3.8 is genuinely the closest thing to Anthropic’s Fable 5 that anyone has shipped. On text it sits fifth, and on agentic coding the DeepSWE gap is wide.

Price is where Qwen 3.8 makes its strongest case. It delivers near-frontier results at a rate well below what closed Western flagships charge, which for high-volume workloads can matter more than the last few leaderboard points.

What a Sparse Mixture-of-Experts Model Means

The 2.4 trillion figure sounds enormous, and it is, but no request ever uses all of it. A sparse Mixture-of-Experts design splits the model into many smaller expert networks and a routing layer sends each token to only the few most relevant ones. The rest of the model stays idle for that token.

With the active count now public, the picture is concrete. Qwen 3.8 activates about 95 billion of its 2.4 trillion parameters per token, roughly 4% of the model. That is why Alibaba can price a multi-trillion-parameter system at $2 per million input tokens; you are paying for the compute that actually runs, not the parameters sitting on disk.

This also explains the open-weight problem. A 4% activation ratio keeps inference affordable on Alibaba’s infrastructure, but downloading the model still means storing all 2.4 trillion parameters. Self-hosting Qwen3.8-Max will need datacenter-scale hardware, which is exactly why the smaller Qwen3.8-27B checkpoint matters for most teams.

What About Qwen 4?

Qwen 3.8 may not sit at the top of Alibaba’s roadmap for long. Leaked roadmap chatter, surfaced by outlets including Geeky Gadgets, points to Qwen 4.0 arriving around September 2026. Treat that timeline as rumour, because Alibaba has not officially confirmed a Qwen 4 date or even its existence.

The most repeated claim is that Qwen 4 will push into 3D coding and design, with uses across gaming, architecture, and virtual reality. Two stealth test models, reportedly codenamed “Caleb” and “Terrania Alpha,” have been floated as early Qwen 4 experiments. All of this traces back to a single leak chain with no benchmarks and no official source, so file it under “interesting if true” and wait for Alibaba to speak.

Should You Use Qwen 3.8?

For developers, the calculus changed at general availability. A model that ranks second on vision and fifth on text at $2 / $6 is a serious cost argument for document processing, visual reasoning, and long-context work. The 1-million-token window paired with $0.25 cached reads makes repeated passes over large corpora unusually cheap.

Two reasons to hold off. If your workload is agentic coding, the DeepSWE 1.1 gap against Fable 5 is real and you should benchmark before switching. If you are an enterprise, the data-residency questions around routing work through a Chinese provider need answering first. The Qwen family is worth tracking either way, including its role powering Apple Intelligence in China.

If you would rather compare leading models without juggling separate logins and API keys, an app like Fello AI routes your prompts to current top models from one place. On a Mac that includes a native Qwen desktop client routed through US-based infrastructure, which answers part of the data-residency question above. For the wider field, our roundup of the best AI models keeps the rankings current.

The Bottom Line

Qwen 3.8 shipped, and it mostly delivered. The 2.4-trillion-parameter scale is real, the 95 billion active figure explains the aggressive pricing, and a #2 vision ranking behind only Fable 5 is a legitimate result rather than a press release. At $2 / $6 it is the strongest price-to-capability argument in the frontier tier right now.

Just keep the claim and the evidence separate. “Second only to Fable 5” is true on vision and wrong on text, where Qwen 3.8 sits fifth. The open weights did ship in mid-August, but the free download is a text-only version of the API model, so check which one a benchmark or provider is actually quoting.

FAQ