Qwen3.8-Max: Alibaba's Open-Weight Coding Flagship

Alibaba announced Qwen3.8-Max on Monday, August 3, 2026, and it is the largest model the Qwen team has ever shipped: 2.4 trillion parameters, a 1 million-token context window, and a deliberate focus on long-horizon coding and agentic work. The headline for builders is not the parameter count. It is that Alibaba says it will open-weight a Max-class model for the first time, with the weights due the following week alongside a smaller Qwen3.8-27B. A frontier-scale coding model you can eventually run yourself changes the math for anyone building on top of large models.

What Qwen3.8-Max actually is

Qwen3.8-Max is a sparse Mixture-of-Experts model. Of the 2.4 trillion total parameters, only a fraction activate per token, which is what keeps inference costs in a workable range for a model this size. It ships today through Alibaba Cloud's Model Studio APIs and through QwenWork, Alibaba's workplace agent platform, so you can call it right now without waiting for the open-weights drop.

According to the South China Morning Post, Alibaba is positioning the model around real-world task completion rather than chat: recreating working software applications from screenshots, running multi-week coding jobs, and driving agentic workflows that span hundreds of steps. The official Qwen announcement frames it as a flagship built for "coding, real-world tasks, in-depth research, and tackling long-horizon challenges."

The benchmarks and the 10-day coding run

Qwen's published figures put the model at the front of the open pack on developer tasks. The team reports SWE-bench Verified results in the high 80s and a Terminal-Bench 2.1 score in the high 60s, and independent testers on the AINews digest noted it lands within a couple of points of the best closed coding models on terminal tasks. Benchmarks are a starting point, not proof, but the pattern is consistent: this is a coding-first model, not a general assistant that also writes code.

The demonstration that got the most attention was an autonomous coding run. Alibaba showed the model working from an empty folder to a production application over more than ten days of unattended operation, with the full commit history, over 500 commits, published publicly so anyone can inspect the trace. Long-horizon autonomy is exactly the capability that separates a code-completion tool from an agent you can hand a project to, and it is the dimension where most models still break down after a few hours.

Agentic and Computer-Use Benchmarks

The clearest signal is OSWorld-Verified, the benchmark that scores computer-use agents driving a real desktop. Qwen3.8-Max posts 86.1, ahead of Fable 5, GPT-5.6 Sol Max, and Gemini 3.1 Pro. It also jumps on FrontierSWE from the predecessor's 40.7 to 73.5, a near-doubling that reflects a real step up in autonomous software engineering. The independent Latent Space AINews breakdown tracks the same pattern across coding and cowork tasks.

The takeaway is narrow but important: on the agentic and computer-use axis, Qwen3.8-Max is now the model to beat. On single-shot reasoning it is competitive rather than dominant. If your workload is a chatbot, the gap is small. If your workload is an agent that opens apps, edits files, and runs commands across a long session, the gap is the whole point.

How it compares to Kimi K3 and DeepSeek V4

Qwen3.8-Max lands in the middle of a remarkable three-week stretch for open-weight coding models. Moonshot's Kimi K3 (2.8T) set the bar in July, and DeepSeek shipped V4-Flash on July 31. Qwen3.8-Max is not the biggest of the three, but it is the first Max-class Qwen release to promise open weights, which matters for teams that need to self-host.

The competitive story is that the gap between open and closed coding models has narrowed to a few benchmark points. For a full breakdown of where the closed frontier sits, see our coverage of Claude Opus 5. If your priority is cost rather than a single top model, open-weight routing approaches like the one in our Echo model router writeup are increasingly viable now that several 2T-class models are open.

What creators and builders can do with it

The practical unlock is agentic coding you can point at a real project. Through the Model Studio API, you can wire Qwen3.8-Max into a coding agent (Cline, an OpenAI-compatible harness, or your own loop) at $2 per million input tokens and $6 per million output tokens, with cached input at roughly $0.25 per million. The 1 million-token context means you can load an entire mid-size codebase, its docs, and a long task brief into a single run instead of chunking and re-summarizing.

Three concrete workflows are worth trying this week. First, screenshot-to-app: hand the model a screenshot of an interface and ask it to rebuild a working front end, which Alibaba highlighted as a first-class use case. Second, overnight refactors: give it a scoped, well-tested module and let a long-horizon run migrate it, then review the commit trace in the morning. Third, research-plus-build tasks where the model reads a paper or spec and produces a runnable implementation. If your team already ships open image models, this is the same lab behind our Qwen-Image-Flash coverage, so a single vendor can now cover both your image and coding stacks.

The open-weights question to watch

The open-weights release is the part to verify before you build a self-hosting plan around it. Alibaba has said the weights for both Qwen3.8-Max and Qwen3.8-27B will land the week of the announcement, and they are expected on the Qwen HuggingFace organization. What is not yet settled is the license. Early reports suggest usage terms may carry regional restrictions, and the exact license text was not public at announcement time. Until the weights and license are actually posted, treat "open weights next week" as a plan, not a shipped fact, and keep the API path as your near-term integration.

Update: The Open Weights Are Now Live

The release lands as two repositories: a full-precision checkpoint and an FP8 quantized checkpoint that is already the more downloaded of the pair. Both are public, both ship complete weights spread across 213 safetensors shards, and the FP8 build had cleared several thousand downloads within hours of going live. The weights carry a custom qwen3.8-max license rather than a standard Apache or MIT grant, so teams should read the terms before building on it commercially.

The architecture is the interesting part. Qwen3.8-Max is a 512-expert MoE that routes to 10 experts plus 1 shared expert per token, with a hidden dimension of 8192. Alibaba pairs a Gated DeltaNet layer stack with periodic Gated Attention layers, a hybrid design meant to keep long-context inference affordable. Native context is 262,144 tokens, extensible to roughly 1,010,000. One detail builders should note: the published checkpoint is a text-only model that requires thinking mode for every interaction, even though the hosted Max API accepts image and video input. The open weights are the reasoning core, not the full multimodal stack.

How to Run It, and Who Actually Can

The honest constraint: 2.4 trillion parameters is enormous, and almost nobody runs this on a single machine. The practical paths are narrower than the download button suggests.

1. Serve the FP8 checkpoint on a multi-GPU node. The FP8 build roughly halves the memory footprint versus full precision and is the intended production target. You still need a serious multi-GPU server, but it brings the model into range for a well-equipped team.

2. Rent capacity rather than buy it. Because the weights are open, inference providers can host them. For most builders the near-term route is calling Qwen3.8-Max through a hosted endpoint while keeping the option to self-host later, rather than provisioning 200-plus GPUs on day one.

3. Wait for the smaller companion. Alibaba has signaled a dense 27-billion-parameter Qwen3.8 checkpoint alongside the Max. That model, not the 2.4T monster, is what most people will actually run locally. As of this release it is not yet on the Qwen Hugging Face org, so it remains a watch item.

Fine-tuning is the other reason open weights matter here. With a closed API you adapt behavior through prompting alone; with the weights in hand you can run supervised fine-tuning or preference tuning on the FP8 or full-precision checkpoint and specialize it for a narrow domain, a house coding style, or a private tool schema. That is out of reach on rented API access, and it is where self-hosting a model this capable starts to pay for the infrastructure it demands.

The AINews roundup tracks both checkpoints and the surrounding open-weight race, which now includes Kimi K3, DeepSeek's latest, and GLM as the other serious contenders shipping downloadable frontier models.

Frequently asked questions

Is Qwen3.8-Max open source right now?

No. As of the August 3 announcement, the model is available only through Alibaba Cloud's Model Studio API and QwenWork. Alibaba has promised to release open weights the following week, but they were not public at launch.

How much does the Qwen3.8-Max API cost?

Reported pricing is $2 per million input tokens and $6 per million output tokens, with cached input around $0.25 per million. That undercuts most closed frontier coding models while offering a 1 million-token context window.

How does it compare to Kimi K3?

Kimi K3 is larger at 2.8 trillion parameters and posts the highest open-weight benchmark scores so far. Qwen3.8-Max is close on coding tasks and differentiates on long-horizon autonomy, with demonstrated multi-week unattended coding runs.

What is Qwen3.8-27B?

It is a much smaller 27-billion-parameter model that Alibaba is open-weighting alongside the flagship. It targets teams that want a Qwen model they can run on a single node or modest GPU setup rather than a 2.4T-parameter system.

Can I use it as a coding agent today?

Yes. Call it through the Model Studio API from an agent harness like Cline or any OpenAI-compatible client. The 1 million-token context and long-horizon behavior make it suited to whole-repo tasks rather than single-file completions.

What should I verify before self-hosting?

Wait for the actual weights and the published license. Confirm the license permits your use case and region, check the real hardware requirements for a 2.4T sparse MoE, and benchmark it on your own tasks before committing a pipeline to it.