· via Vercel blog
Alibaba's Qwen4-Max reasoning model with 1M-token context arrives on Vercel AI Gateway
Alibaba's Qwen4-Max, the first model in the Qwen4 series, is now available through Vercel's AI Gateway, adding a 1M-token reasoning model with native vision for coding agents and API users.

Qwen4-Max, the newest large reasoning model from Alibaba, is now reachable through Vercel's AI Gateway. According to Vercel's changelog, it is the first release in the Qwen4 series: a reasoning model with native vision support and a context window spanning one million tokens, aimed at coding, professional document work and long-running agent jobs.
What the model can do
Vercel highlights Qwen4-Max's ability to operate real creative and professional software. The model can construct 3D scenes in Blender and build playable games in Unity, Unreal Engine and Godot, and, notably, run what it builds, using visual feedback to correct camera angles, animation glitches and lighting problems.
For fields such as finance, law and healthcare, Vercel says the model can orchestrate many MCP tool calls, process extensive collections of documents and images, and produce spreadsheets, slide decks and reports that cite their sources. The million-token window is what makes that document-heavy work plausible in fewer passes.
Built for long-horizon agents
The most distinctive claim concerns behavior on extended tasks. Rather than working invisibly until a job finishes, Qwen4-Max produces a workable first version early and refines it across successive iterations, so a human can review progress and redirect the work before it completes. Alibaba is releasing a Live Delivery plugin as open source to support this pattern.
Vercel also reports comparisons against earlier Qwen Max models: the new version finishes the same tasks while consuming fewer tokens, and its generated code contains fewer injection, memory-safety and access-control vulnerabilities. The changelog does not include benchmark figures, so these remain vendor claims for now.
How to use it
The model is exposed under the identifier alibaba/qwen4-max. To route it into coding agents such as Claude Code, Codex or Cursor, install the latest Vercel CLI and run npx vercel ai-gateway setup, then select the model inside the agent. It can also be tried in Vercel's model playground.
The AI Gateway itself offers a single API for calling many models, together with usage and cost tracking, retries, failover and other performance tuning that Vercel says delivers uptime above what individual providers manage. It also includes custom reporting, Zero Data Retention support, spending budgets per API key and routing rules. Pricing mirrors the underlying provider rates, with no markup and no platform fee on inference, including on requests made through a customer's own API keys under BYOK.
Why it matters
The launch matters on several levels. First, gateway-style access keeps lowering the friction of switching between frontier models: a team already routing Claude or GPT models through Vercel can add a million-token reasoning model with a configuration change rather than a new integration. Second, the model's approach to long tasks, shipping a rough version early and iterating in the open, targets one of the most common frustrations with agents, which is discovering only at the end that they went the wrong way. Third, the emphasis on lower token consumption and safer default code speaks to the two costs that dominate production agent deployments: spend and security exposure.
The caveats are the usual ones. Capability, efficiency and security claims all come from Alibaba and Vercel rather than independent evaluation, and there is no shortage of vendors making similar promises about long context and agentic reliability. Even so, the arrival of another very-long-context reasoning model on a mainstream developer platform sharpens competition in a segment that has largely been defined by Western labs, and gives developers an easy way to test those claims against their own workloads.
- #ai
- #alibaba
- #vercel
- #ai-gateway
- #llm