Skip to content

A Qwen VL model on Bedrock Mantle refuses image input, because it inherits the text-only default class #195

Description

@JGoutin

Found by the #175 research sweep, in released code.

qwen.qwen3-vl-235b-a22b-instruct binds to the generic Bedrock Mantle chat class, which declares INPUT_MODALITIES = ("TEXT",) (stdapi/models/chat/_mantle/_default.py:102). The model-layer modality check at stdapi/models/__init__.py:3434 therefore refuses image input for it.

That is wrong on two independent counts: the model's runtime twin publishes TEXT, IMAGE, and the committed probe record for this model shows image_input: accepted.

This is the published-capability defect #175 anticipated — a Mantle matcher that fails to bind does not raise, so the model still answers text requests and no existing test fails. It is only visible as a capability the catalogue advertises differently from what the gateway will accept.

Proposed shape: a new _mantle/qwen_vl.py family rather than widening open_weight, since the distinguishing declaration is a modality set, not a prefix.

Filed separately from #175, which is test-only: this is a production fix and wants its own fix: commit.

Ref: stdapi/models/chat/_mantle/, tests/test_aws_bedrock_mantle.py::TestMantleModelClassResolution

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingmodelAI/LLM model specific

    Type

    No type

    Projects

    • Status
      In review

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions