If your team is still paying GPT-4o rates for long-document and multimodal workloads, the numbers are about to look embarrassing.

Repository-scale AI was too expensive for most teams to run seriously

Coordinating coding agents across an entire codebase, or processing hours of video alongside dense legal documents, has been a cost and context-window problem. Most hosted models either truncate the input or charge enough to make the workflow impractical at volume.

Two million tokens per minute, text and video in, structured answers out

Alibaba Qwen Releases Qwen3.8 accepts text, images, and video as input and returns text, accessed today via an OpenAI-compatible API requiring only a base-URL and model-ID swap. The context window sits at 1 million tokens, with a maximum input of 991K tokens and output of 131K tokens. A thinking mode is available with a 262K reasoning-token budget, and the rate limit is 15K requests per minute — numbers that suggest Alibaba built this for production throughput, not demos.

Legal, engineering, and media teams are the first to feel the gap close

• Senior software engineers running repo-scale coding agents who need a model that holds the entire codebase in context without chunking hacks.
• Legal and financial analysts reviewing long-form documents who need structured extraction across hundreds of pages in a single pass.
• Media and e-commerce operators indexing long video content who need timestamped, searchable output without manual tagging.

A companion checkpoint, Qwen3 235B-A22B at 27B activated parameters, targets teams that want on-premise deployment on standard multi-GPU hardware without a datacenter build-out.

The price gap between Alibaba and OpenAI just became a procurement conversation

At $2.00 per million input tokens and $6.00 per million output tokens, with implicit cache reads at $0.25, Qwen3 235B-A22B lands meaningfully below GPT-4o pricing for high-volume document and multimodal workloads. Open weights arrive next week, which means the self-hosted cost floor drops further — and every enterprise AI vendor now has a pricing problem to answer.

What this model actually lets you run today

• Run a long-video indexing pipeline against 60-minute recordings in one API call.
• Drop a 900K-token legal contract into context and extract structured clause summaries.
• Build a multi-step research assistant with a 262K reasoning budget and no chunking.
• Swap your current OpenAI integration to Qwen3 235B-A22B with a single endpoint change.

Pricing is $2.00 per million input tokens and $6.00 per million output tokens via the hosted API; open-weights self-hosting costs depend on your infrastructure.

The 2.4T total parameter count means the full open-weight checkpoint is a multi-node datacenter artifact — Alibaba has not disclosed activated-parameter count, so serving cost for self-hosting cannot yet be fully modeled.

GPT-4o and Gemini 1.5 Pro both offer comparable context windows, but neither ships open weights at this parameter scale, which matters the moment your legal or compliance team asks where the data goes.

The open-weights race at frontier scale just changed what CFOs will approve

When 2.4T-parameter open-weight models become available next week, the conversation inside every enterprise AI procurement meeting shifts from “which API vendor” to “do we self-host.” We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.