Overview
Qwen3.8-Max is Alibaba Qwen's new flagship model, officially released on August 3, 2026. The release moves the Max line to a 2.4 trillion parameter model with 95 billion active parameters, while keeping Qwen focused on practical agent work rather than only short benchmark prompts. Qwen describes it as the most capable model in the Qwen family to date, with stronger coding, cowork, research, long-horizon task execution, and multimodal agent capability.
The launch is also important for model access strategy. Qwen says this is the first Qwen-Max-class model that will receive open weights, but those weights were not downloadable at launch. The official plan is to release them during the week after August 3, 2026, so teams should treat hosted QwenCloud access as the immediate production route and open-weight deployment as a near-term follow-up.
What's New
2.4T Max-Scale Model
Qwen3.8-Max succeeds the Qwen3.7-Max flagship line with 2.4 trillion total parameters and 95 billion active parameters. The model builds on the architectural foundation of Qwen3.5, but targets harder end-to-end workflows: codebases, research reproduction, competition pipelines, professional document work, and agent systems that need to keep state across many tool calls.
For ToolWorthy readers comparing AI coding tools, the practical signal is that Qwen is positioning Qwen3.8-Max as a coding agent model, not just a chat model with coding answers.
Autonomous Coding and Research Workflows
The official release examples focus on long-running coding tasks. Qwen reports a 10+ day autonomous coding run for a self-evolving CLI project, a research-paper reproduction and improvement task, and an online multimodal-dialogue competition workflow. These examples emphasize planning, code generation, test execution, feedback loops, and leaderboard-style iteration.
The more relevant takeaway for builders is not that every task should run unattended for days. It is that Qwen3.8-Max is designed for agent harnesses where the model has to plan, edit, run tools, inspect failures, and continue without constant human prompting.
Stronger Cowork and Professional Task Coverage
Qwen's launch article also tests the model against professional workflows across law, finance, design, manufacturing, healthcare-style demos, sports analytics, menu planning, and structural modeling. The common pattern is broad work execution: ingesting messy documents or briefs, producing a structured deliverable, and maintaining enough task context to complete the workflow in one session.
That makes Qwen3.8-Max relevant for teams building AI agent systems that need dependable outputs, not only conversational assistance.
Long-Horizon and Closed-Loop Optimization
Qwen highlights long-horizon tasks such as autonomous chip-design optimization and continuous learning in operational settings. These examples show the model using feedback loops instead of a fixed single-shot plan: propose an action, run a tool or simulation, inspect the result, and update the next step.
This is a useful distinction from ordinary long-context summarization. Qwen3.8-Max is meant to use context, tools, and iterative evaluation together.
Multimodal Agent Inputs
The official assistant and agent configuration examples list text and image input support. Qwen also describes the model as improving multimodal agents, including workflows involving screenshots, images, manuscripts, and video-like work inputs. For production planning, treat text and image support as the directly documented model configuration, and verify any video-specific workflow limits in current QwenCloud documentation before building around them.
Benchmark Notes
Qwen published a full benchmark table comparing Qwen3.8-Max with Opus 4.8, Fable 5, GPT-5.6 Sol, and Qwen3.7-Max. Selected Qwen3.8-Max results include:
| Area | Benchmark | Qwen3.8-Max |
|---|---|---|
| Coding agent | Terminal Bench 2.1 | 86.6 |
| Coding agent | SWE-bench Pro | 67.7 |
| Coding agent | PaperBench | 93.0 |
| Coding agent | QwenSWEBench | 80.7 |
| General agent | CoWorkBench | 74.8 |
| General agent | WorkSpaceBench | 67.7 |
| General agent | WideSearch | 81.9 |
| General capability | IFBench | 82.8 |
| General capability | GPQA Diamond | 92.6 |
| Long context | MRCR v2 256K, 8-needle | 92.9 |
These are vendor-published numbers, so they are best read as launch-positioning data until independent evaluations appear. Still, the table is directionally useful: Qwen3.8-Max is strongest in the release narrative where agentic coding, professional work, and long-context execution overlap.
Availability & Access
Qwen3.8-Max is available through QwenCloud at launch. The official article provides API examples for QwenCloud's OpenAI-compatible endpoint and Anthropic-compatible endpoint, plus integration examples for Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw.
The model identifier shown in Qwen's examples is qwen3.8-max. Agent configuration examples list a 1,000,000 token context window and 65,536 maximum output tokens. The same examples list reasoning support and text-plus-image inputs.
Open-weight access is not live on August 3, 2026. Qwen states that the Qwen-Max-class open weights are planned for release the following week. Teams that need self-hosting should wait for the official weights, license, model card, and hardware guidance before committing deployment architecture.
Pricing & Plans
Qwen's launch article does not publish a standalone token-pricing table for Qwen3.8-Max. Immediate hosted access is through QwenCloud / Alibaba Cloud Model Studio, so production costs should be verified on the current Model Studio pricing page or inside the QwenCloud console before launch.
For now, the safest planning assumption is:
- Qwen Chat / Qwen Studio: useful for hands-on evaluation where access is available.
- QwenCloud API: the production route at launch, billed according to Model Studio account terms and current regional pricing.
- Open weights: planned for the week after August 3, 2026, but not yet a usable self-hosting option at launch.
Best For
- Engineering teams building long-running coding agents that edit, test, and recover from failed runs.
- Research teams reproducing papers, running experiments, and iterating on results.
- Product teams building AI chatbot or cowork assistants that need structured deliverables rather than short answers.
- Enterprises testing agent workflows across legal review, financial research, manufacturing analysis, or document-heavy operations.
- Open-model teams waiting for a Max-class Qwen release they can evaluate for self-hosting after the weights ship.
FAQ
Is Qwen3.8-Max officially released?
Yes. Qwen officially released Qwen3.8-Max on August 3, 2026, through the Qwen blog and made hosted access available via QwenCloud.
Is Qwen3.8-Max open source?
Not on launch day. Qwen says Qwen3.8-Max will be the first Qwen-Max-class model with open weights, and that the weights are planned for release the week after August 3, 2026. Until those files and license details are live, treat it as hosted-access first.
What is the Qwen3.8-Max model ID?
The official API and agent examples use qwen3.8-max.
Does Qwen3.8-Max support a 1M context window?
Yes. Qwen's official agent configuration examples list contextWindow as 1,000,000 and maxTokens as 65,536.
Does Qwen3.8-Max support images?
Yes. Official configuration examples list text and image input support. The launch article also describes Qwen3.8-Max as stronger for multimodal agents, but teams should verify any video-specific input limits in current QwenCloud documentation.
How much does Qwen3.8-Max cost?
Qwen did not publish a dedicated Qwen3.8-Max pricing table in the launch article. For production API use, check the current Alibaba Cloud Model Studio or QwenCloud pricing page for your account region and token tier.
