MindWorkAI/AI-Studio
1
0 Fork 0
mirror of https://github.com/MindWorkAI/AI-Studio.git synced 2026-10-06 22:09:40 +00:00
Code Issues Packages Projects Releases Wiki Activity

Recognize the models which hold a connection open instead of chatting

Browse Source
This commit is contained in:
Thorsten Sommer committed 2026-09-19 16:58:05 +02:00
1 parent f7350bf86b
commit 987ccf527d
5 files changed
+62 -8
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

app/Tests/Models/Corpus/CapabilitySnapshot.txt
+1 -1
View File
@@ -64,7 +64,7 @@ FIREWORKS | accounts/fireworks/models/qwen3-235b-a22b | CHAT_COMPLETION_API, FUN
FIREWORKS | whisper-v3 | SPEECH_INPUT, TEXT_OUTPUT | TRANSCRIPTION | (unknown) | (unknown) | (unknown)
GOOGLE | gemini-1.0-pro-vision | CHAT_COMPLETION_API, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
GOOGLE | gemini-2.0-flash | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | (unknown) | 3600 per request | PROVIDER_API countTokens
GOOGLE | gemini-2.0-flash-live-001 | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, SPEECH_INPUT, SPEECH_OUTPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | (unknown) | (unknown) | (unknown)
GOOGLE | gemini-2.0-flash-live-001 | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, SPEECH_INPUT, SPEECH_OUTPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | REALTIME | (unknown) | (unknown) | (unknown)
GOOGLE | gemini-2.5-flash | ALWAYS_REASONING, AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | 1048576 | 3600 per request | PROVIDER_API countTokens
GOOGLE | gemini-2.5-flash-image | CHAT_COMPLETION_API, IMAGE_OUTPUT, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | IMAGE_GENERATION | (unknown) | (unknown) | (unknown)
GOOGLE | gemini-2.5-flash-lite | AUDIO_INPUT, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, OPTIONAL_REASONING, SPEECH_INPUT, TEXT_INPUT, TEXT_OUTPUT, VIDEO_INPUT | CHAT | 1048576 | 3600 per request | PROVIDER_API countTokens
app/Tests/Models/Corpus/ModelKindCorpus.cs
+37
View File
@@ -92,6 +92,10 @@ public static class ModelKindCorpus
new(GOOGLE, "imagen-4.0-generate-001", IMAGE_GENERATION, AnsweredTodayAs: CHAT, Reason: "The markers never knew the name; the family ported in the Google step states it. Nobody noticed because the Google provider shows only names beginning with gemini."),
new(ALIBABA_CLOUD, "qwen-image-edit", IMAGE_GENERATION),
// Google ships one image model under a codename instead of a description. Nothing in it
// says drawing, and the provider only ever saw it because its catalog was read whole.
new(GOOGLE, "nano-banana-pro-preview", IMAGE_GENERATION),
];
/// <summary>
@@ -172,6 +176,21 @@ public static class ModelKindCorpus
//
new(ALIBABA_CLOUD, "qwen-tts-realtime", REALTIME),
new(ALIBABA_CLOUD, "qwen3-asr-flash-realtime", 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
// and all four stood among the models to talk to -- which none of them can be. The last
// one translates between two people speaking, which is as far from a chat as it gets.
//
new(GOOGLE, "gemini-3.8-live", REALTIME),
new(GOOGLE, "gemini-3.8-live-extended-thinking", REALTIME),
new(GOOGLE, "gemini-3.1-flash-live-preview", REALTIME),
new(GOOGLE, "gemini-3.5-live-translate-preview", REALTIME),
// Google's experimental music model, which is recognized already: it carries the other
// word, and this holds that the two rules do not fall out with each other.
new(GOOGLE, "lyria-realtime-exp", REALTIME),
];
/// <summary>
@@ -188,6 +207,11 @@ public static class ModelKindCorpus
private static readonly ModelKindExample[] TEXT_COMPLETION_ENTRIES =
[
new(HELMHOLTZ, "text-davinci-003", TEXT_COMPLETION),
// The model Mistral serves for filling a gap in a file. Its name does not begin with the
// word the provider used to sort its list by, so it stood among the chat models -- next to
// Codestral, which does the same job and was kept out by that very word.
new(MISTRAL, "mistral-code-fim-latest", TEXT_COMPLETION),
new(OPEN_AI, "babbage-002", TEXT_COMPLETION),
new(OPEN_AI, "gpt-3.5-turbo-instruct", TEXT_COMPLETION),
@@ -279,6 +303,19 @@ public static class ModelKindCorpus
new(PERPLEXITY, "sonar-reasoning-pro", CHAT),
new(PERPLEXITY, "sonar-deep-research", CHAT),
//
// The counter-sample to the three rules above, taken off the same three catalogs. Every
// one of these carries a word which now means something -- code, live, image -- without
// being what that word says, and every one of them has to stay a chat model.
//
new(MISTRAL, "mistral-code-latest", CHAT),
new(MISTRAL, "mistral-vibe-cli-latest", CHAT),
new(MISTRAL, "zai-glm-latest", CHAT),
new(GOOGLE, "gemma-4-31b-it", CHAT),
new(GOOGLE, "gemini-3.8-flash", CHAT),
new(X, "grok-4.6", CHAT),
new(X, "grok-build-0.1", CHAT),
//
// Three which were questioned while testing and stay all the same. Grok Build is the coding
// model behind the xAI CLI and answers like any other Grok. The Groq compound systems are