Alibaba just shipped the biggest model it’s ever released. Qwen3.8-Max went from preview to full launch in about two weeks – and it’s dragging the rest of the open-weight world along with it.
Qwen3.8-Max is Alibaba’s flagship AI model, built on 2.4 trillion total parameters with roughly 95 billion active per token via a sparse Mixture-of-Experts architecture. It handles text, images, video, and documents, supports a 1 million token context window, and is available now through Alibaba Cloud’s API – with open weights following shortly after launch.
What Is Qwen3.8-Max?
Qwen3.8-Max is the newest flagship in Alibaba’s Qwen family, built on the architectural foundation of Qwen3.5. The model contains 2.4 trillion total parameters, with approximately 95 billion parameters activated during inference, and it supports a context window of up to 1 million tokens. That’s enough to hold entire codebases, long documents, or hours of video in a single request.
Here’s the thing that actually matters, though – and it’s easy to miss if you just skim the “2.4 trillion” headline. Total parameter count is not the same as usable compute. Qwen3.8-Max runs on a sparse Mixture-of-Experts (MoE) design, which means only a slice of that 2.4 trillion fires on any given request. Alibaba paired this with a hybrid attention mechanism, cutting computational cost and latency compared to dense models of similar scale.
Timeline: From Preview to Full Release
The rollout happened in three stages, and the gaps between them tell you something about how Alibaba ships models now.
- July 19, 2026 – Alibaba unveiled Qwen3.8-Max-Preview at the World AI Conference in Shanghai, claiming performance “second only to Claude Fable 5.” No benchmark table, no model card, no license, no open-weights date.
- August 3–4, 2026 – Alibaba made Qwen3.8-Max broadly available and published a full benchmark table, confirming open weights would ship the following week.
- August 12–13, 2026 – Two checkpoints landed on Hugging Face: Qwen3.8-27B under Apache 2.0, and the full 2.4T Qwen3.8-2.4T-A95B checkpoint under a custom license.
That preview-first pattern is worth flagging if you’re the kind of person who reads a launch tweet and assumes you can deploy the same day. You couldn’t – not for three weeks.
Specs at a Glance
| Spec | Detail |
| Total parameters | 2.4 trillion |
| Active parameters per token | ~95 billion |
| Architecture | Sparse Mixture-of-Experts, hybrid attention |
| Context window | Up to 1 million tokens |
| Max output tokens | 128,000 |
| Modalities | Text, image, video, document input |
| Access | Alibaba Cloud Model Studio (QwenCloud), OpenAI/DashScope-compatible API |
| Open weights | Qwen3.8-27B (Apache 2.0), Qwen3.8-2.4T-A95B (custom license) |
The open-weight version omits a few things the hosted API has – image input and a non-thinking mode aren’t included in the release. So if you’re planning to self-host, check which capabilities you actually need before you commit rack space to it.
How Does It Perform? The Benchmarks
Alibaba’s own framing was “second only to Fable 5.” The actual numbers are more of a mixed bag – strong in some places, clearly behind in others.
Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Claude Fable 5 (both at 84.6), but behind GPT-5.6 Sol at 88.8. On SWE-bench Pro it posts 67.7 against Fable 5’s 80.0, and on FrontierSWE it lands at 73.5 versus Fable 5’s 88.8. It does lead on PaperBench at 93.0 and IFBench at 82.8, and GPQA Diamond comes in at 92.6 – a small bump over its predecessor.
Where it actually pulls ahead is agentic and multimodal work. It tops most vision benchmarks, including OSWorld-Verified at 86.1, Parametric CAD Bench at 91.5, and OmniDocBench 1.5 at 92.1. And compared to Qwen3.7-Max, the jump on long-horizon agentic tasks is large: DeepSWE 1.1 moves from 21.6 to 56.6, FrontierSWE goes from 40.7 to 73.5, and JobBench climbs from 31.3 to 53.4.
One caveat worth keeping: the multimodal comparison table benchmarks against Qwen3.7-Plus rather than Qwen3.7-Max, which flatters the generation-over-generation gap. Read the “biggest leap yet” claims with that in mind.
What About OMI CLI?
Alongside the model, Alibaba released something unusual: an autonomous coding tool the model built for itself. OMI CLI is a coding harness that Qwen3.8-Max built and evolved over roughly 16 days of self-directed operation, starting from an empty code repository – and Alibaba published it publicly. It’s a small case study in what “agentic” actually means in practice, beyond the marketing copy.
How Does It Compare to the Competition?
Qwen3.8-Max didn’t launch into an empty field. It arrived days after Moonshot AI’s Kimi K3 – a 2.8 trillion parameter open-weight model that’s technically larger, launched the same week. Size alone puts Qwen3.8-Max second among publicly known open models, behind Kimi K3.
Positioning-wise, Alibaba is aiming it squarely at closed frontier labs. It’s pitched as a direct competitor to closed models like Anthropic’s Fable line and OpenAI’s GPT-5.6 Soul, especially on coding and long-running agentic tasks. Alibaba also said the model ranks fifth in Text Arena and second in Vision Arena – leaderboard positions worth checking independently rather than taking at face value, since self-reported arena rank shifts fast.
How to Access Qwen3.8-Max
You’ve got two real paths in:
- Hosted API – available now through Alibaba Cloud Model Studio, and it’s OpenAI- and DashScope-compatible, so switching over is mostly a base-URL and model-ID change if you’re already on an OpenAI-style integration.
- Self-hosted open weights – Qwen3.8-27B under Apache 2.0 fits on standard on-premise GPU hardware, while the full 2.4T Qwen3.8-2.4T-A95B checkpoint is a multi-node datacenter artifact – not something you’re running on a single box.
If you’re evaluating this for a product, start with the API. The full-scale checkpoint isn’t a weekend-project download.
FAQ
Is Qwen3.8-Max open source?
Partially. Alibaba released the weights as Qwen3.8-2.4T-A95B in mid-August 2026, though the open-weight version drops some features from the cloud model, like image input. A smaller companion model, Qwen3.8-27B, ships under the fully permissive Apache 2.0 license.
How many active parameters does Qwen3.8-Max use?
Around 95 billion parameters are activated during inference, out of 2.4 trillion total, thanks to its Mixture-of-Experts design.
Is Qwen3.8-Max better than GPT-5.6 or Claude Fable 5?
It depends on the task. It beats both on some coding benchmarks like Terminal-Bench 2.1, but trails Fable 5 on SWE-bench Pro and FrontierSWE. It’s strongest on vision and long-horizon agentic tasks, not raw reasoning.
What is the context window for Qwen3.8-Max?
Up to 1 million tokens in, with 128,000 tokens of output.
