Ask the endpoint for transcription models before falling back to the documented list

app/MindWork AI Studio/Provider/Fireworks/ProviderFireworks.cs

@@ -93,6 +93,16 @@ public class ProviderFireworks() : BaseProvider(LLMProviders.FIREWORKS, new Uri(
     }
     
     /// <inheritdoc />
+    /// <remarks>
+    /// The one transcription list which stays a plain list, where GWDG and Mistral ask their
+    /// endpoint first. There is nothing to ask here: HasModelLoadingCapability is false and every
+    /// other method above answers with nothing, which is why a chat model at Fireworks has to be
+    /// typed in by hand. A list is all there is.
+    ///
+    /// The commented-out entry is no oversight either. The documentation names Whisper v3 Turbo,
+    /// and trying it does not work -- which is worth keeping written down, so that nobody adds it
+    /// back and finds out the same way again.
+    /// </remarks>
     public override Task<ModelLoadResult> GetTranscriptionModels(string? apiKeyProvisional = null, CancellationToken token = default)
     {
         // Source: https://docs.fireworks.ai/api-reference/audio-transcriptions#param-model

app/MindWork AI Studio/Provider/GWDG/ProviderGWDG.cs

@@ -18,6 +18,12 @@ public sealed class ProviderGWDG() : BaseProvider(LLMProviders.GWDG, new Uri("ht
         new("qwen3-embedding-4b", "Qwen3 Embedding 4B"),
     ];
 
+    // Source: https://docs.hpc.gwdg.de/services/saia/index.html#voice-to-text
+    private static readonly Model[] KNOWN_TRANSCRIPTION_MODELS =
+    [
+        new("whisper-large-v2", "Whisper v2 Large"),
+    ];
+
     #region Implementation of IProvider
 
     /// <inheritdoc />
@@ -123,13 +129,27 @@ public sealed class ProviderGWDG() : BaseProvider(LLMProviders.GWDG, new Uri("ht
     }
     
     /// <inheritdoc />
-    public override Task<ModelLoadResult> GetTranscriptionModels(string? apiKeyProvisional = null, CancellationToken token = default)
+    /// <remarks>
+    /// Built the same way as the embedding models above, and for the same reason: SAIA answers the
+    /// models endpoint with its chat models only, so this comes back empty and the documented list
+    /// stands in. Asking first costs nothing and means a speech model appearing in that answer one
+    /// day shows up on its own, rather than waiting for somebody to notice and edit this file. A
+    /// failed request is passed on unchanged, so a wrong API key stays visible as such.
+    /// </remarks>
+    public override async Task<ModelLoadResult> GetTranscriptionModels(string? apiKeyProvisional = null, CancellationToken token = default)
     {
-        // Source: https://docs.hpc.gwdg.de/services/saia/index.html#voice-to-text
-        return Task.FromResult(ModelLoadResult.FromModels(
-        [
-            new Model("whisper-large-v2", "Whisper v2 Large"),
-        ]));
+        var result = await this.LoadModels(SecretStoreType.TRANSCRIPTION_PROVIDER, apiKeyProvisional, token);
+        if (!result.Success)
+            return result;
+
+        var transcriptionModels = result.Models.Where(model => model.IsTranscriptionModel(this.Provider)).ToList();
+        if (transcriptionModels.Count is 0)
+            return ModelLoadResult.FromModels(KNOWN_TRANSCRIPTION_MODELS);
+
+        return result with
+        {
+            Models = [..transcriptionModels]
+        };
     }
     
     #endregion

app/MindWork AI Studio/Provider/Mistral/ProviderMistral.cs

@@ -124,13 +124,16 @@ public sealed class ProviderMistral() : BaseProvider(LLMProviders.MISTRAL, new U
     }
     
     /// <inheritdoc />
-    public override Task<ModelLoadResult> GetTranscriptionModels(string? apiKeyProvisional = null, CancellationToken token = default)
+    public override async Task<ModelLoadResult> GetTranscriptionModels(string? apiKeyProvisional = null, CancellationToken token = default)
     {
-        // Source: https://docs.mistral.ai/capabilities/audio_transcription
-        return Task.FromResult(ModelLoadResult.FromModels(
-        [
-            new Provider.Model("voxtral-mini-latest", "Voxtral Mini Latest"),
-        ]));
+        var modelResponse = await this.LoadModelList(SecretStoreType.TRANSCRIPTION_PROVIDER, apiKeyProvisional, token);
+        if (!modelResponse.Success)
+            return modelResponse;
+
+        return modelResponse with
+        {
+            Models = [..modelResponse.Models.Where(n => n.IsTranscriptionModel(this.Provider))]
+        };
     }
     
     #endregion

app/Tests/Models/Corpus/ModelKindCorpus.cs

@@ -122,6 +122,13 @@ public static class ModelKindCorpus
         new(SELF_HOSTED, "wav2vec2-large-xlsr-53", TRANSCRIPTION),
         new(MISTRAL, "voxtral-mini-latest", TRANSCRIPTION),
 
+        // The rest of what Mistral actually serves, off its own catalog. The app offered the one
+        // name above alone, because that is the one the documentation names; these three were there
+        // the whole time. Both sizes come as a rolling name and as a dated snapshot.
+        new(MISTRAL, "voxtral-small-latest", TRANSCRIPTION),
+        new(MISTRAL, "voxtral-mini-2602", TRANSCRIPTION),
+        new(MISTRAL, "voxtral-small-2507", TRANSCRIPTION),
+
         // NVIDIA's other speech line, once plain and once as the hub names it. The second one is
         // what makes the rule a segment worth keeping: the organization comes off before any rule
         // sees the name, so what is left has to carry the word on its own.
@@ -154,6 +161,11 @@ public static class ModelKindCorpus
         new(SELF_HOSTED, "xtts-v2", SPEECH_SYNTHESIS),
 
         new(ALIBABA_CLOUD, "qwen-tts", SPEECH_SYNTHESIS),
+
+        // The Voxtral which speaks instead of listening. It stands here to hold the other half of
+        // Mistral's transcription list: the catalog is asked now, so what is not a transcription
+        // model has to be kept out by what it is, not by the list having been short.
+        new(MISTRAL, "voxtral-mini-tts-latest", SPEECH_SYNTHESIS),
     ];
 
     /// <summary>
@@ -177,6 +189,12 @@ public static class ModelKindCorpus
         new(ALIBABA_CLOUD, "qwen-tts-realtime", REALTIME),
         new(ALIBABA_CLOUD, "qwen3-asr-flash-realtime", REALTIME),
 
+        // The other two of Mistral's Voxtral line. The second one carries "transcribe" and stays
+        // out of the transcription list all the same: whatever it does, it does over a connection
+        // the app cannot open, and that is what the rank on the realtime rule is for.
+        new(MISTRAL, "voxtral-mini-realtime-latest", REALTIME),
+        new(MISTRAL, "voxtral-mini-transcribe-realtime-2602", REALTIME),
+
         //
         // Google's whole two-way line, taken off its catalog rather than written from memory. It
         // uses the other word for the thing OpenAI calls realtime, so none of these was recognized