Minimax-m3 is not recognizing images in SaaS

[juanmichelini]

See

notice that claude-sonnet-4-6 see images fine:

Notice that images were visualized fine in benchmark https://results.eval.all-hands.dev/swebenchmultimodal/litellm_proxy-minimax-MiniMax-M3/26986845327/results.tar.gz

[juanmichelini]

I just inspected the swebench‑multimodal results tarball linked in this issue (the litellm_proxy-minimax-MiniMax-M3/26986845327 run) and the multimodal pipeline on the agent side looks healthy, and the model does appear to be using the images:

1. Images are being attached correctly

• 64/68 conversation tarballs have at least one image attached to the first user MessageEvent as a multimodal content part:

  {"type": "text", "text": "I have access to a code repository … <issue_description> …"},
  {"type": "image", "image_urls": ["https://user-images.githubusercontent.com/1595739/…png"]}

• The other 4 are chartjs__Chart.js-{8868,9199,9678,9766}, whose source issues legitimately don't have an attached screenshot (they only link to CodePen / jsFiddle / chartjs docs samples). Not a pipeline issue.
• Image counts per conversation (1–6, mostly 1) match what swebench‑multimodal ships. All URLs are valid HTTPS user-images.githubusercontent.com links.

2. The model is reading the images, not just receiving them
Concrete visual claims in agent ActionEvent.thought / messages that could only come from looking at the image:

• processing__p5.js-5555 (WebGL alpha‑blend bug, red square on a black canvas over a blue page):

  "The pixel value is rgba(128, 0, 0, 255), which is red blended with black (the canvas background), not blue (the body background). … After fix: Red blends with black canvas background (rgba(128, 0, 0, 255)) — matching the 2D renderer behavior as requested in the issue."

• processing__p5.js-5915 (per‑vertex color gradient in WEBGL):

  "Now let me visually verify by looking at the screenshot … The first line is working with proper gradient! The first line goes from white (203, 203, 255) to blue (50, 50, 255), with a smooth gradient." and "Lines 2 and 3 are not at those exact pixel coordinates (my sample points are off for the diagonals)".

• processing__p5.js-6069 (additive blend bug):

  "…blue, second point's color) instead of the additive blend result (R=255, G=0, B=255)."

And explicit "Image N" / "screenshot" references that match the system prompt's Image 1, Image 2, … markers (so the system‑injected image labels are reaching the model too):

• processing__p5.js-5305: "Image 1 - Some errors to be added …"
• Automattic__wp-calypso-22621: "Looking at the image …"
• Automattic__wp-calypso-23915, 25778, 34435: "looking at the screenshot"
• diegomura__react-pdf-1541: "The screenshots show a small bar in the middle vs a full‑width bar."
• processing__p5.js-3680: "the image, the C shape is visible. The image in the issue shows a different rendering."

So in this benchmark the image path through the litellm proxy → M3 is working and the model is reasoning about visual content. The bug in the SaaS UI screenshot looks specific to the SaaS frontend (the issue shows a "file selected, name preview" but the model reports "no images") rather than the agent/eval pipeline itself.

[all-hands-bot]

Minimax-m3 image recognition failure in SaaS shares the same root cause (SDK _supports_vision detection failing for models behind LiteLLM proxy) as #3768 (Deepseek-v4-pro), #3767 (Kimi-k2.6), and #3495 (step-3.7-flash). However, the maintainer intentionally files per-model issues, as confirmed by the bot triage comment on #3768 noting these are 'distinct' despite shared mechanism. Not a duplicate, but worth cross-linking.

Found 3 possible duplicate issues:

1. #3768 — Deepseek-v4-pro cannot see images in SaaS
2. #3767 — Kimi-k2.6 is not being recognized as a vision model in the SDK
3. #3495 — step-3.7-flash: vision support not detected at runtime (LiteLLM proxy /v1/model/info lookup returns None)

These may not be exact duplicates, but the scope appears to overlap enough that keeping discussion in one place may be more useful.

This comment was created by an AI assistant (OpenHands) on behalf of the repository maintainer.