Summary

Ollama-compatible clients (Open WebUI, Enchanted, ollama-grid-search, etc.) rely on the capabilities list and details.{parameter_size,quantization_level,families} fields returned by /api/tags and /api/show to filter which models are eligible for a given task. LocalAI's compat layer was leaving these empty, so clients silently rejected embedding/rerank models for chat and refused to surface them as embedding models.

This PR wires up the existing signals already present in ModelConfig:

• capabilities:
  • embedding — FLAG_EMBEDDINGS
  • completion — FLAG_CHAT / FLAG_COMPLETION
  • vision — explicit KnownUsecases bit, MMProj, multimodal template, or backend-reported media marker
  • tools — auto-detected ToolFormatMarkers, JSONRegexMatch/ResponseRegex, XMLFormat[Preset], grammar triggers, or schema type
  • thinking — ReasoningConfig with reasoning not disabled, or explicit thinking markers
  • insert — presence of a completion template
• details.families — populated from the backend name (was previously empty)
• details.parameter_size / details.quantization_level — parsed from the GGUF filename via conservative regex (won't match Qwen3 as 3B)
• model_info.general.architecture / model_info.general.context_length — exposed in the new model_info map on ShowResponse

A small gotcha worth flagging: cfg.HasUsecases(FLAG_VISION) can't be used directly because GuessUsecases has no FLAG_VISION case and ends up returning true for every chat model. hasVisionSupport() instead checks KnownUsecases bits explicitly plus the multimodal projector / template / media-marker signals. This is documented inline.

Fixes #9760.

Test plan

Built with TDD using Ginkgo/Gomega — 11 new specs in capabilities_test.go + models_test.go:

• DescribeTable for modelCapabilities: embedding-only, plain chat, vision-capable chat, thinking, tools via auto-detected markers, tools via explicit JSON regex, pure backend-only model (rerank) returns no caps
• modelDetailsFromModelConfig populates format=gguf, family, families, parses Qwen3-4B-Instruct-Q4_K_M.gguf → parameter_size=4B, quantization_level=Q4_K_M
• End-to-end /api/show JSON tests using a real ModelConfigLoader populated from a temp YAML file — round-trips through ShowModelEndpoint and verifies Capabilities + Details.{ParameterSize,QuantizationLevel,Format,Families}

$ go test ./core/http/endpoints/ollama/...
ok  	github.com/mudler/LocalAI/core/http/endpoints/ollama	0.037s
$ go test ./core/schema/... ./core/config/...
ok  	github.com/mudler/LocalAI/core/schema	0.006s
ok  	github.com/mudler/LocalAI/core/config	0.349s
ok  	github.com/mudler/LocalAI/core/config/meta	0.017s

Manual verification for the reviewer

To reproduce the issue from #9760 against the patched build, with any embedding model configured:

$ curl -s http://localhost:8080/api/show -d '{"name":"qwen3-embedding"}' | jq '.capabilities, .details'
[
  "embedding"
]
{
  "format": "gguf",
  "family": "llama-cpp",
  "families": ["llama-cpp"],
  "parameter_size": "4B",
  "quantization_level": "Q4_K_M"
}

Open WebUI / Enchanted / ollama-grid-search should now correctly surface embedding models in their embedding-model picker, and stop offering them as chat models.