Qwen3.8-Max
#9 in LLM models
« Alibaba's very large, multimodal, open-source model. Performs particularly well on coding, research, or long, autonomous tasks. Supports a context size of up to 1 million tokens»
Qwen3.8-Max: 2.4 trillion parameters, and for the first time you can download them
Qwen3.8-Max is Alibaba's biggest multimodal language model to date, aimed at coding and long-running agent work. It packs a 2.4-trillion-parameter mixture-of-experts design with 95 billion active per query and a one-million-token context window on the API. In a first for the Max line, the weights sit on Hugging Face for anyone to grab, while the hosted API bills by the token. Engineering teams get OpenAI- and Anthropic-compatible endpoints from day one.
Pros
- One-million-token context window on the API
- First Max-class Qwen with downloadable weights
- Endpoints compatible with OpenAI and Anthropic formats
- Very strong on long-run autonomous agent work
Cons
- Open-weight version accepts text only
- Far too heavy for desktop GPUs
- Custom license, stricter than Apache 2.0
Qwen3.8-Max codes for sixteen days straight without supervision
Qwen3.8-Max takes a software project from an empty folder to a finished deliverable over days of continuous work, with no human in the loop. Alibaba's headline demonstration ran for roughly sixteen days on its own, racking up 265 commits and 127 pull requests in a public GitHub repository.
The model also watches what it does. Images and video feed a correction loop during execution rather than serving as plain input. On OSWorld-Verified, a test of how well an agent drives a real desktop, Alibaba reports 86.1, ahead of every proprietary model it measured. Vendor figures, so run your own checks, yet independent trackers already place it near the very top for reasoning.
Qwen3.8-Max tested against Kimi K3 on real coding work
Putting the model's image and video input to the test
Open weights on Hugging Face, with a license worth reading twice
The full mixture-of-experts weights can be downloaded from Hugging Face and ModelScope under the name Qwen3.8-2.4T-A95B, a first for a Max-class model at Alibaba. An FP8 variant ships alongside, and the bespoke license includes a revenue-share clause for the largest commercial deployments.
One caveat before you clear disk space (forum threads did grumble about this at length), the open checkpoint is no carbon copy of the hosted service. The table below sums up the gap.
| Feature | Hosted API | Open weights |
|---|---|---|
| Accepted input | Text, images, video | Text only |
| Context window | 1 million tokens | 262K native, 1M via YaRN |
| Cost | Billed per token | Free download |
| License | Alibaba service terms | Custom license with revenue share |
Plugging Qwen3.8-Max into your current tools without rewriting anything
Qwen3.8-Max speaks both the OpenAI and Anthropic API formats, so existing coding assistants such as Claude Code, Codex and Qwen Code work with it right away. You point your SDK at the new base URL, swap the model id, and you are off.
On the bill, count $2 per million input tokens and $6 per million output, with implicit caching at $0.25, below the list price of the previous Qwen3.7-Max. Rates in this market move fast, so Alibaba Cloud's official grid gets the final word.
- QwenCloud and Alibaba Cloud Model Studio for direct access
- OpenRouter and Vercel's AI Gateway for a two-minute trial
- Hugging Face and ModelScope for self-hosting
Frequently asked questions
Is Qwen3.8-Max free?
Downloading the weights from Hugging Face costs nothing, though you supply the server underneath. The hosted API bills per million tokens, and Alibaba also sells credit-based subscription plans. For a quick trial without an Alibaba account, OpenRouter added the model one day after launch and remains the fastest route.
Can I use Qwen3.8-Max commercially?
Yes through the API, like any hosted service. The open weights ship under a bespoke license that departs from the Apache 2.0 terms of earlier Qwen releases and includes a revenue-share clause for very large commercial deployments, with the threshold still being finalized. Read the license file in the repository before shipping anything.
Qwen3.8-Max or Kimi K3, which should I pick?
Kimi K3 carries 2.8 trillion parameters and independently audited scores, but its output tokens cost two and a half times as much. Qwen3.8-Max works out far cheaper and behaved as the tidier agent in the one third-party head-to-head published so far, finishing with zero failed tool calls.
What hardware does Qwen3.8-Max require for self-hosting?
Server-grade, multi-GPU hardware is the entry ticket. At 2.4 trillion total parameters, the model sits in the same deployment class as Kimi K3 and GLM-5.2, well away from the local 30B tier. An FP8 checkpoint trims memory use, and a 27-billion-parameter sibling was announced for single-GPU machines.
Verdict: Handing an entire repository to an agent that keeps grinding for days is where Qwen3.8-Max earns its keep. Engineering teams after frontier-level coding with a self-hosting option, served through endpoints their tools already understand, will feel right at home.
★ Featured AI Tools ★
AI Alternatives for
Qwen3.8-Max
Paid
Paid
Paid
Paid
Paid
