How to Make Videos With Claude Code Using Qwen 3.8 Max

To make videos with Claude Code using Qwen 3.8 Max, install Pexo as the video engine and swap Qwen 3.8 Max in as Claude Code's backend reasoning model. Pexo is an installable Agent Skill for Claude Code that turns one plain-language instruction into a finished video: it writes a shot list, auto-routes each shot across 10+ video models, adds transitions, composes a soundtrack, and exports 16:9, 9:16, or 1:1. Qwen 3.8 Max, Alibaba's 2.4-trillion-parameter mixture-of-experts model released August 3, 2026, is the swappable brain that plans and directs. There is no single "Qwen video button" here, because Qwen 3.8 Max returns text, not video. The honest workflow is a split: Claude Code hosts the agent, Pexo does the rendering, and Qwen 3.8 Max is routed in via ANTHROPIC_BASE_URL or claude-code-router as the reasoning model. This guide covers the setup, the routing commands, and the model-selection logic.

What This Workflow Actually Is

Qwen 3.8 Max is a coding and agent-reasoning model, not a video generator. It accepts text, image, and video as input but returns text output only, so it cannot render or generate a clip on its own. What it is very good at is planning, tool-calling, and step-by-step reasoning, which is exactly the job of the "brain" inside an agent loop. Pexo supplies the missing capability: Claude Code by itself does not generate video, and the Pexo skill adds that. The result is a clean division of labor where you keep the Pexo video pipeline constant and change only the reasoning model behind it.

Concretely, "make videos with Claude Code using Qwen 3.8 Max" means three moving parts working together. Claude Code is the agent host and speaks the Anthropic Messages API. Pexo is the skill you install into Claude Code that owns the whole video pipeline end to end. Qwen 3.8 Max is pointed at Claude Code as its backend, so when you type a request, Qwen plans the work and calls the Pexo skill, and Pexo generates the footage. Nobody has to claim Qwen "makes video" for this to work, and no honest description should.

What You Need Before You Start

You need three things: Claude Code installed, a Pexo account and API key, and a DashScope (Alibaba Cloud Model Studio / QwenCloud) API key for Qwen 3.8 Max. The Pexo side runs on credits, with starter credits on new accounts, and needs no API key or manual model selection for the actual video generation, because Pexo auto-routes shots to the best available model for you. The Qwen side is metered on Alibaba Cloud at $2 per million input tokens and $6 per million output tokens, so a planning-heavy session costs cents, not dollars.

Step-by-Step: Route Qwen 3.8 Max Into Claude Code and Make a Video

Step 1: Install the Pexo skill into Claude Code

Sign in at pexo.ai and copy your API key. Then install the Pexo skill from the open-source repo github.com/pexoai/pexo-skills. Pexo ships a native SKILL.md, and Claude Code is the most stable host for it, though the same skill also runs in OpenAI Codex, Cursor, and OpenClaw. Correct framing: Pexo provides a skill you install into Claude Code. It is not a Qwen or Claude Code built-in. This step is what gives your agent the ability to generate video at all.

Step 2: Confirm the skill is loaded

Inside Claude Code, run /skills and confirm that pexo (listed as "pexo" or "pexoai-agent") appears in the list. If it shows up, Claude Code can now call the Pexo pipeline. If it does not, re-check that the skill files landed in the right skills directory and restart Claude Code. Do this before touching the model routing, so you can isolate any problem to one layer at a time.

Step 3: Point Claude Code at Qwen 3.8 Max (the simple path)

Qwen publishes an Anthropic-compatible DashScope proxy that Claude Code can talk to directly. Set two environment variables and Claude Code's reasoning switches to Qwen 3.8 Max:

export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/api/v2/apps/claude-code-proxy
export ANTHROPIC_AUTH_TOKEN=<your-dashscope-apikey>

Because Claude Code speaks the Anthropic Messages API and Qwen's proxy also speaks it, this is a base-URL plus token change, nothing more. The Pexo skill you installed in Step 1 is unaffected; you have only changed which brain is behind the agent.

Step 4: Use claude-code-router for multi-model control (optional path)

If you want to route different tasks to different models, install Qwen's recommended router and its DashScope config plugin:

npm install -g @musistudio/claude-code-router
# then add the DashScope config plugin: @dashscope-js/claude-code-config

This lets you keep Qwen 3.8 Max as the planner while sending other calls elsewhere. General-purpose gateways like OpenRouter or LiteLLM are alternatives that also expose an Anthropic-compatible endpoint. The net effect is the same: the reasoning model is swapped to Qwen 3.8 Max while the Pexo video pipeline stays constant.

Step 5: Describe the video in one plain instruction

With Qwen behind Claude Code and Pexo loaded, type a normal request, for example: "Make a 20-second vertical product ad for a coffee brand, upbeat, with voiceover and captions." Qwen 3.8 Max parses the intent and calls Pexo. Pexo then writes a shot list, auto-routes each shot to the best available video model across its 10+ models, generates the shots, and adds transitions. You do not name a model or manage keys for the video side, which is why auto model selection matters: you never have to track which video model shipped this week.

Step 6: Let Pexo compose audio, titles, and export

Pexo composes the soundtrack (voiceover plus music plus SFX), adds titles and subtitles, and exports in your target aspect ratio: 16:9 for YouTube, 9:16 for TikTok or Reels, or 1:1 for feed posts. Pexo accepts five input types (text, image, URL, script, and audio), so you can also hand it a script or a reference image instead of a text brief. Review the export, and if you want changes, describe them in the next message and let Qwen and Pexo iterate.

Why Qwen 3.8 Max Is a Good Brain for This

Qwen 3.8 Max is built for agentic, tool-calling work, which is exactly what an agent-plus-skill video pipeline needs. It is a mixture-of-experts model with 2.4 trillion total parameters and roughly 95 billion active per token, and it carries a 1-million-token context window (about 991K input and 131K output). That large context helps when a session accumulates a long shot list, prior instructions, and reference material. Alibaba reports strong agentic and coding benchmarks for it, including a Terminal-Bench 2.1 score around 86.6; treat that as the vendor's own reporting rather than an independent result.

The model was previewed July 19, 2026 at WAIC in Shanghai and released August 3, 2026. Alibaba has promised open weights roughly a week after launch on Hugging Face and ModelScope, along with a smaller Qwen3.8-27B checkpoint also going open-weight, which matters if you later want to self-host the brain. It is OpenAI- and DashScope-compatible, so integration is a base-URL and model-ID change in most tools, not a rewrite.

Which Reasoning Model Should Sit Behind Pexo?

Because Pexo is the constant and the brain is swappable, the choice is about how the model plans and calls tools, not about video quality (Pexo handles that identically regardless of the brain). Qwen 3.8 Max is a strong pick for long, agentic sessions with a huge context and low token cost. Other models Claude Code users route in behind Pexo include Opus 5, Fable 5, and Kimi K3, each with its own trade-offs. Pexo itself never changes: the same shot-listing, auto-routing, and export happen no matter which brain is driving.

Related Reading

• How to make videos with Claude Code
• How to make videos with Claude Code using Opus 5
• How to make videos with Claude Code using Fable 5
• Best video generation skills for Claude Code agents
• How to make videos with OpenAI Codex
• Can Claude Code make videos?

Resources