The shipped deepseek-v4-flash catalog entry drops images for a backend that supports them

[luckystation-cmd]

Summary

The provider has retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp model names: requests using either name are served by DeepSeek-V4.1-Flash and billed at the Flash price. V4.1-Flash supports vision. The shipped catalog still declares deepseek-v4-flash as text-only, so a user who selects that entry and attaches an image has the image replaced by deterministic text even though the backend that answers the request can see it.

Evidence

Provider page https://api-docs.deepseek.com/quick_start/pricing, model details:

• deepseek-flash → DeepSeek-V4.1-Flash, 1M context, Vision ✓
• footnote (1): "Use deepseek-flash as the model name. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price."

Repository catalog, packages/llm/llm-deepseek/src/index.ts:92-122:

{ id: 'deepseek-flash', name: 'DeepSeek-V41-Flash', inputModalities: ['text', 'image'], … },
{ id: 'deepseek-v4-flash', name: 'DeepSeek-V4-Flash',
  description: 'Fast, efficient, and economical; suited to focused, routine, or parallel tasks.', … },   // no inputModalities → ['text']
{ id: 'deepseek-v4-pro', name: 'DeepSeek-V4-Pro', description: …, … },                                    // text-only
{ id: 'deepseek-v4-flash-vision-exp', name: 'DeepSeek-V4-Flash-Vision-Exp', inputModalities: ['text', 'image'], … },

packages/llm/llm-deepseek/README.md:49 states the intent: "Omitted models advertises the text- and image-capable deepseek-flash and deepseek-v4-flash-vision-exp alongside the text-only deepseek-v4-flash and deepseek-v4-pro."

Impact

1. Silent capability loss. LlmRuntime.subject substitutes an image the route cannot accept with deterministic text (projectImagesForTextModel → textOnlyImageText), and packages/llm/llm-deepseek/src/request-pricing.ts:79 prices the same substitution. A user on deepseek-v4-flash gets a text placeholder where the served model would have read the image. The failure is invisible: nothing errors, the model simply answers from a description.
2. Two picker entries for one backend. deepseek-v4-flash, deepseek-v4-flash-vision-exp, and deepseek-flash are now the same model at the same price, so the picker offers a meaningless choice and the deepseek-v4-flash description ("Fast, efficient, and economical") implies a cheaper tier that does not exist.
3. Misleading affordance. The vision-capable entry is the one named after a retired experimental variant, so the entry that looks safest to pick is the one that loses images.

Proposal

Pick whichever the maintainers prefer; both are small:

• Correct the stale entry: give deepseek-v4-flash inputModalities: ['text', 'image'] plus the image budgets it is served with, and drop or alias the now-redundant deepseek-v4-flash-vision-exp. This restores capability without changing what a user types.
• Drop the retired names from DEFAULT_MODELS so the picker offers deepseek-flash (vision) and deepseek-v4-pro (text-only) only. The ids still pass through to the wire, so an explicit models list keeps working for anyone who needs them.

Either way, deepseek-v4-flash's description should stop implying a cheaper tier, and the README's paragraph at line 49 should match the new catalog.

Acceptance criteria

• Selecting deepseek-v4-flash (or whatever the catalog keeps) and attaching an image sends the image, not a text substitution — or the id is no longer offered as a selectable entry.
• The catalog no longer advertises a price/quality tier that the provider does not bill.
• packages/llm/llm-deepseek/README.md:49 and the catalog agree.