DeepSeek V4 Pro is text-only planning

Same attachment rule as GLM: no stills on the wire. Cost-efficient when the brief is already words. Not a video model.

DeepSeek V4 Pro is a chat brain on the agent. It plans, talks, and calls tools. It does not draw the clip. Send an image part and the provider 404s — the same constraint as the GLM 5 family. Pin it when the brief is already a script, a scene list, a credit budget. Name Kling 3 Turbo or Seedance 2.5 in the same message if a file has to come back.

The open-weight story is DeepSeek V4. This page is the attachment rule and the two-catalog split.

Planner, not a catalog row

Picking the chat model behind the agent is the rule. Chat models live on GET /api/v1/agentic/chat-models. Generation models live on GET /api/v1/agentic/models and on the public catalog. Swap the first and the agent plans differently. The MP4 still comes from whichever video row the generate tool was told to use.

DeepSeek V4 Pro is on the built-in chat roster as a text-only id. Send the id from /chat-models, not the marketing name. If you save it as a custom OpenRouter row, the slug shape is org/model — the public listing is deepseek/deepseek-v4-pro. Duplicate ids are dropped.

The documented default is runpod/kimi-k2.7-code. DeepSeek V4 Pro is not the default. Neither is GLM 5.3. Treat /chat-models as live.

Chat tokens and generations bill on two meters. Changing to DeepSeek changes the reasoner line. It does not change what a named renderer costs. There is no free chat allowance. It is not on /models. If the picker mixed “DeepSeek” with a video family, you picked the wrong catalog.

The still 404s

From the built-in registry, text-only brains include the GLM 5 family, MiniMax M2.7, Qwen3 Max Thinking, the MiMo line, DeepSeek V4 Pro, and the RunPod-hosted Kimi ids (including the default). Send an image to those and the image part 404s.

Practical rule:

• Brief is already words. DeepSeek V4 Pro is a fair pin for a long plan-and-call loop.
• Brief is a pack shot, a screenshot, a frame. Pick a vision brain — Grok 4.6 (x-ai/grok-4.6), Grok 4.3, Gemini 3.1 Pro, Claude, Qwen3.5 397B — or put the visual in text and skip the attachment. Grok 4.6 can see the still; GLM 5.3 cannot is the same split. DeepSeek sits on the GLM side of it.
• Video or audio attachments. Those are pre-processed through Gemini for non-Gemini brains. You do not have to pick Gemini just to describe a clip. You do have to pick vision if the image is the brief.

Do not “fix” the 404 by asking DeepSeek to imagine the product. That is a generate you did not pin. Upload the still to a vision chat model, or generate from a locked file with image-to-video after the plan is text.

GLM 5.3 is a planner, not a renderer is the sibling pin (z-ai/glm-5.3, custom OpenRouter, always-on reasoning). DeepSeek is the cost-efficient built-in when you want that attachment rule without switching to a custom Z.ai id.

Pin the brain, name the renderer

On POST /api/v1/agentic/chat and /chat/stream send chat_model as the id from /chat-models. Send chat_model_explicit: true when the value is a choice, not a leftover default.

On the OpenRouter path, fast-eligible sub-agents (generation, slideshow, social, web editor) may run on Claude Haiku 4.5 unless that flag is set. Workflow scene-directing and movie planning keep the brain you picked. The default RunPod Kimi is not swapped for Haiku.

A lasting preference can store “plan with DeepSeek V4 Pro.” It still does not pin Kling.

Prompt shape:

Use this chat model for planning. Do not change the video model. Generate with Kling 3 Turbo, 5s, 9:16, locked-off bottle on wet stone. If you need allowed durations, call the schema tool, then generate. Do not substitute a different renderer.

If you omit the second sentence, a newly cheap brain may “help” by routing to a different family. You will blame DeepSeek for a look it did not render. When you are exploring planners, keep the renderer named. When you are exploring looks, keep the brain stable.

Nine chat models to pin is the install list. DeepSeek is item eight: text-only, same attachment rule as GLM, not a video model.

Cost-efficient when the brief is already words

The job this brain wins is volume of text in, tools out: scene lists, credit quotes, “call the schema then generate,” rewriting VO lines, routing a week of briefs that never needed a jpeg on the wire. That is why it is the cheap planner when the campaign already exists as words.

It is the wrong pin when the brief is the still. A pack shot in the composer is a vision job. DeepSeek will 404 or, if you stripped the image, plan from a description you typed — which is a different, lossy brief.

It is also the wrong pin if you wanted a video model called DeepSeek. There isn’t one. The generate row is Kling, Seedance, Veo, or Grok Imagine Video — named in the brief, billed on the generation meter.

Skills are a third door. npx skills add AI-XLabs-Innovation/versely-skills installs generate / slideshow / movie / ugc / music / social / analytics / pipeline into a host agent. That does not change the in-app chat model. Pin DeepSeek in /agent. Install skills in Claude Code.

FAQ

Is DeepSeek V4 Pro the default Versely chat model?

No. Treat /chat-models as live. The documented default at the time of the two-catalog write-up is runpod/kimi-k2.7-code. DeepSeek V4 Pro is a text-only built-in you send or save. GLM 5.3 is a different, custom OpenRouter id.

Why did my photo attachment fail?

DeepSeek V4 Pro is text-only. Image parts 404. Use Grok 4.6, Gemini 3.1 Pro, or Claude when the still is the brief, or describe the still in text and skip the attachment.

Will pinning DeepSeek change my next clip’s look?

Only if the new brain picks a different generate model than the last one did. Pin the renderer in the brief. The chat model changes planning, tool use, and the text reply.

Can I use DeepSeek V4 Pro to generate video directly?

No. It is not a video model. It can call generate tools. The file comes from the catalog row you named — or from a router pick you should have named. Two catalogs.