Skip to content

models: deepseek-v4.1-flash-free rejects image input though advertised as image-capable #61842

Closed as duplicate

Description

@aitools1

Summary

Sending an image to opencode/deepseek-v4.1-flash-free fails with [invalid_request_error] image input is not available on this engine, even though the model is advertised as image-capable. The same model accepted screenshots until 2026-10-31, so this looks like a gateway/engine regression.

Environment

  • opencode version: 2.2.5
  • OS: Windows 10.0.26200 (win32 x64)
  • Terminal: Unavailable: TERM_PROGRAM not set (TERM=xterm-256color)
  • Shell: C:\Windows\system32\cmd.exe
  • Install/channel: desktop app bundle (ai.opencode.desktop/resources/opencode-cli.exe), latest
  • Active plugins: [email protected], opencode-dcp, @dietrichgebert/ponytail

Reproduction

  1. Create a small PNG (426 bytes, red background with a blue circle).

  2. Run:

    opencode run -m opencode/deepseek-v4.1-flash-free -f imgtest.png "What shapes and colors are in this image? One short line."
    
  3. Optionally repeat with -m "opencode/deepseek-v4.1-flash-free#max".

Expected Behavior

The model describes the image. models.dev lists opencode/deepseek-v4.1-flash-free as modalities.input: ["text","image"], attachment: true, so image input should be accepted.

Actual Behavior

Error: Upstream request failed: [invalid_request_error] image input is not available on this engine
(request_id: gw-ae0e2e66bcf8-192db7)

Exit code 1, no reply. Same for the #max variant (request_id: gw-add10c6ab910-192efe).

Deterministic: two fresh CLI runs plus six interactive attempts on 2026-11-04 between 23:25 and 23:30 UTC, all rejected. Engine IDs observed: gw-addc59e92621d, gw-addc3cc2c204, gw-adba25bd59c1, gw-ae0e2e66bcf8, gw-add10c6ab910 — all text-only.

Additional Context

Control run with the same file works:

opencode run -m opencode/nemotron-3.5-super-vl-free -f imgtest.png "..."
> A blue circle on a red background.

Other models on the same account:

  • opencode/deepseek-v4.1-flash-vision-exp → Model unavailable
  • opencode/deepseek-v4.1-flash (paid) → Insufficient account funds (unrelated: account balance)

Regression window from local session history (~/.local/share/opencode/opencode.db), session ses_e78708eacffdFd9QXgh11m6r6B, model deepseek-v4.1-flash-free:

  • 2026-10-29 12:08, 12:54, 12:56, 15:08, 15:13, 15:16, 15:20, 15:28, 15:56, 16:08 UTC — image/png attached, assistant replied normally (some replies explicitly describe screenshot content).
  • 2026-10-31 12:36, 12:37, 13:02, 13:15, 13:18, 13:20, 13:25, 13:26, 15:44, 16:06, 16:48 UTC — image/png and image/jpeg attached, normal replies ("На скриншоте видно, что Wildberries…").
  • 2026-11-04 23:25 UTC — first observed image input is not available on this engine.

No config change on my side between those dates; no updates other than routine use.

Workaround: switch to opencode/nemotron-3.5-super-vl-free for image input.

Activity

  1. opencodejr commented on Nov 4, 2026

    @opencodejr
    Contributor

    Thanks for the detailed report, including the request IDs and the dates from your session history.

    I think this is a duplicate of #61827. It reports the same image input is not available on this engine error from the gateway, at the same time (2026-11-04), on 2.2.5, with DeepSeek V4.1 Flash. A commenter there hit it on opencode/deepseek-v4.1-flash-free, the same model you used. The error comes from the gateway or upstream engine, not from the OpenCode client.

    Possible duplicates and related issues, ordered by likelihood:

    Your details would help on #61827: it fails every time with a plain -f attachment, images worked on 2026-10-31, and you have the gw-… request IDs. Please consider adding them there. Until it's resolved, opencode/nemotron-3.5-super-vl-free works for image input, as you found.

    The bot will close this issue as a duplicate in 1 day unless you reply explaining how your problem is different.

  2. aitools1 commented on Nov 4, 2026

    @aitools1
    Author

    Agreed, this is a duplicate of #61827 — no objection to closing it here.

    I moved the extra data points there: #61827 (comment)

    What it adds beyond what's already in #61827:

    • Reproduced with a plain -f attachment of a 426-byte PNG in a fresh session, so no clipboard paste, no history re-send and no MCP tool result is needed to trigger it.
    • Two more engine prefixes hitting the same error: gw-ae0e2e66bcf8, gw-add10c6ab910 (plus gw-addc59e92621d, gw-addc3cc2c204, gw-adba25bd59c1 from the interactive attempts) — different nodes from the gw-adac8a2c2d3d one in go: deepseek-v4.1-flash image rejection bricks sessions ("image input is not available on this engine") #61827.
    • The same model handled PNG/JPEG screenshots normally on 2026-10-29 and 2026-10-31, so the regression is bounded to roughly Nov 1–4, not "always broken".
  3. opencodejr commented on Nov 4, 2026

    @opencodejr
    Contributor

    Thanks for confirming, and for adding the details to #61827. The -f repro in a fresh session, the extra gw-… engine IDs and the Nov 1–4 window should all help maintainers look into the gateway side. Please follow #61827 for updates. The bot will close this issue as a duplicate as planned.

  4. opencodejr commented on Nov 5, 2026

    @opencodejr
    Contributor

    Closing this as a duplicate of #61827, as agreed. Your extra details are now on #61827: the -f repro in a fresh session, the extra gw-… engine IDs and the Nov 1–4 window. Please follow that issue for updates. Thanks again for the thorough report.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions