Repository navigation
[Bug]: BYOK vision-capable model is incorrectly reported as not supporting vision #213
Description
Product or interface
CLI - interactive TUI
Version
0.4.12-dev
Platform
macOS
OS version and architecture
macOS 26.6.2
Issue area
Model output quality
Desktop log upload ID (optional)
No response
Steps to reproduce
- Configure a BYOK provider in minimax-code and select a model
deepseek-v4.1-flashthat accepts image input (vision). - Start a session with that model.
- Take screenshots and tell mcode to read it.
- minimax-code reports that the selected model does not support vision.
Environment details (please confirm or replace before submitting):
- Provider type: Ollama Cloud
- Model: deepseek-v4.1-flash:cloud
- Permission mode: yolo
- Command without keys:
mcode-dev - Image input method: paste image
Expected and actual behavior
Expected: because the configured BYOK model accepts image input, minimax-code should forward the image to the model and process it as a vision request.
Actual: minimax-code reports that the configured model does not support vision and refuses the image input, even though the model itself is vision-capable.
Impact: vision workflows are unusable with BYOK vision models.
Notes: the vision capability check appears to rely on a hard-coded model list or provider metadata instead of the BYOK configuration, so models it does not recognise are treated as text-only.
Redacted error summary
Screenshots
Image 1:
Image 2:
Additional Information
I think there should be a way to configure how the model brought by BYOK sets its Vision capabilities, but I can't find a way to configure it. Is it just a matter of adjusting ~/.minimax/config.yaml?
Before submitting
- I have searched existing issues.
- I have included my version and removed sensitive information.
Activity
smartworldarafath commentedon Sep 19, 2026
Hi @MiniMax-AI maintainers,
I ran into this as well while testing BYOK vision models, investigated the issue, and put together a fix on my fork:
fix/byok-vision-model-support6ec8ad084ea53dffc2786e6c9e931f2714b17306What was happening
model-ref.ts,capabilitiesFromModelConfigwas only checkingmodalities.inputorcapabilities.support_image === true. It wasn't picking upattachment: true, array-form modalities (e.g.['text', 'image']),vision: true, or top-level flags.UserModelInputViewandMcodeProviderModelInputwere missing thecapabilitiesfield, so capability metadata was getting dropped when saving or updating custom providers.local-model-resolver.tswasn't checkingcatalogModel.inputfor image support.mcode provider adddidn't have a--support-imageflag for declaring vision support from the CLI.Changes
capabilitiesFromModelConfiginmodel-ref.tsto recognize vision capabilities across all common formats (attachment: true, array modalities,vision: true, etc.).capabilitiestoUserModelInputViewandMcodeProviderModelInput.service-input.tsandservice-helpers.ts.deriveModelInputinlocal-model-resolver.tsto check catalog model input modalities.--support-imageflag tomcode provider add.model-ref.test.ts,service-input.test.ts, andprovider-application.test.ts.Testing
pnpm typecheckandpnpm buildpass cleanly with no errors.check:source,check:tsconfig).Since PRs are limited to collaborators according to
CONTRIBUTING.md, could someone please take a look and cherry-pick the commit6ec8ad084ea53dffc2786e6c9e931f2714b17306or grant collaborator access?Thanks!