Qwen3.5 drops a seeing agent you can keep

Qwen3.5 is the open-weight drop the month was waiting for. Not another chat model with a vision adapter taped on. A native vision-language agent: see the deck, call the tool, stay on the brief. The flagship is huge on paper and cheap to run, because only a slice of it wakes up per token. Alibaba also shipped a hosted Plus tier the same week, so the two-lane pattern is honest now. Weights if you want custody. An API if you want someone else to hold the cluster.

That is the split we have been using since December, and this month made it obvious for agents, not just stills.

A model that can see a screen and still forget why it clicked is a demo. A model that can see a screen and finish the ticket is a coworker.

The agent that looks at the work

Qwen3.5-397B-A17B is the headline. Apache 2.0. A million-token window if your serving stack can hold it. The pitch is computer use: desktop, mobile, the messy apps a marketer already lives in. We wrote the operational case in Qwen3.5 397B-A17B Drops Open Weights. The useful sentence is shorter. You can download it. You can fine-tune it. You can refuse to send the brand book to a hosted chatbot because "the open one is not ready."

The smaller cuts landed later. 122B, 35B, 27B, then a pocket range that wants to live on a phone. That ladder is the product. A planner on the cluster. A router on the laptop. A function caller on the handset. Do not spend the flagship deciding which tool to open.

Claude Opus 4.6 and Sonnet 4.6 were the closed half of the same stretch. Long context. Agents that stay on a chain. They are very good. They are also a bill. Use them when the brief is ugly and the deadline is mean. Do not use them to rename a column.

Grok 4.20 sold a four-agent architecture like it invented the intern. Cute. The question is still whether the four of them stop. A swarm that cannot halt is not a workforce. It is a group chat with a credit card.

Can you automate it

Yes, if the agent returns a schema you can check. Tool name, arguments, a stop. Qwen3.5 will do that on your metal. Claude will do it on theirs. Pick by where the work has to live, not by a leaderboard screenshot.

The other test is uglier and more useful. Give it a real deck, not a toy. Ask it to extract the offer, the audience, and the three claims legal will fight. If it invents a fourth claim, it is not ready to sit next to a campaign. If it asks a clarifying question and waits, you might have something.

The clip that takes a pile of references

Seedance 2.0 is the hosted video everyone forwarded. Not a prompt and a prayer. Up to nine images, a few seconds of video, a few audio clips, and a text line, into a multi-shot take with native stereo. Character lock that looks like a campaign instead of a slot machine. BytePlus will take the money. CapCut will take the hobbyists.

That is a real workflow change. Mood boards, a face, a voice, a product still. The model has something to hold besides your adjectives. It is still rent. There are no weights. Price the continuity like you price a vendor editor, and write down what happens when the price moves.

Grok Imagine grew up into a ten second 720p clip with audio and an API. Cheap enough to spray. Not consistent enough to trust a face across a cut. Spray is a mood board. A cut is a promise.

Lyria 3 showed up in Gemini and wrote a 32 second bed from a prompt or a picture. Useful for a scratch track. Not a library you own. If the campaign has to live for a year, you still need a person who can license music, or a stem you already paid for.

What a marketer can actually plug in

If the job is a social burst and you do not care who holds the file, Seedance will get you a first assembly. If the job is a brand film that has to live on your machines, you are still on LTX-2 or Hunyuan from last month. This month did not open those weights further. It just made the rented clip look more like a brief.

The brief is the unlock. A hosted model that can take a face, a voice, and a product still is closer to how a creative team already works than a text box that asks you to describe "cinematic lighting" for the fifteenth time. Closer is not owned. Do not build the archive on a vendor's goodwill.

Stills in the background

xAI shipped Imagine stills built for throughput. Fine. Throughput without a kit is how you get twelve slightly different logos. Qwen's image stack from last winter is still the open edit path. Nobody this month beat "change the hat, keep the room."

If a new stills model cannot survive a revision, it is a toy, no matter how fast it is. Speed is how demos win. Kits are how campaigns ship.

An API is not a stack you own

Seedance. Claude. Grok. Lyria. Gemini. All callable. None of them send you a checkpoint. That is fine if you know it. It is expensive if you pretend it is a stack.

An API is a good thing. It means you can automate the call. It does not mean you can keep the model. Those are different sentences. Say both out loud before you pitch the board a "platform."

Leave March with one seeing agent

One open agent that can see. Qwen3.5 if you can serve it, a smaller cut if you cannot. One hosted video path only if you have accepted the invoice. Keep LTX-2 for the work that cannot leave the building.

Put the agent on a short tool list. Put the rented clip on a budget line. Put the local film on a machine that still works when the vendor has a bad week.

GroundSlate is where the local half sits on a Mac. The agent calls tools. The footage stays on the desk. The rented half is a line item, not an identity.