mirror of
https://github.com/MindWorkAI/AI-Studio.git
synced 2026-10-06 15:29:40 +00:00
Port the Mistral open weights and the GLM family
This commit is contained in:
9 files changed
+215
-7
No files matched your search
@@ -0,0 +1,31 @@
|
||||
using AIStudio.Provider;
|
||||
|
||||
using static AIStudio.Provider.Capability;
|
||||
|
||||
namespace AIStudio.Models.Microsoft;
|
||||
|
||||
/// <summary>
|
||||
/// E5, the embedding models built on somebody else's weights.
|
||||
/// </summary>
|
||||
/// <remarks>
|
||||
/// "e5-mistral-7b-instruct" is what made this a family of its own. It is an embedding model, and it
|
||||
/// carries the name of the model it was trained from, so the Mistral rules answer for it and tell
|
||||
/// it that it chats and calls functions. Saying which name means what it says is cheaper than
|
||||
/// teaching every family whose weights somebody built an embedder from.
|
||||
///
|
||||
/// The E5 part is the whole statement: the rest of the name says nothing about what the model does.
|
||||
/// </remarks>
|
||||
public sealed class E5Family : ModelFamily
|
||||
{
|
||||
/// <inheritdoc />
|
||||
public override ModelVendor Vendor => ModelVendor.MICROSOFT;
|
||||
|
||||
/// <inheritdoc />
|
||||
public override ModelSource Source => new("https://huggingface.co/intfloat/e5-mistral-7b-instruct", new DateOnly(2026, 9, 11), "The app lists this under IProvider.GetEmbeddingModels, which is where the statement that it embeds comes from.");
|
||||
|
||||
/// <inheritdoc />
|
||||
protected override void Declare(ModelFamilyBuilder builder) =>
|
||||
builder.Rule("e5").AsSegment()
|
||||
.Capabilities(TEXT_INPUT | EMBEDDING)
|
||||
.Kind(ModelKind.EMBEDDING);
|
||||
}
|
||||
@@ -0,0 +1,37 @@
|
||||
using static AIStudio.Provider.Capability;
|
||||
|
||||
namespace AIStudio.Models.Mistral;
|
||||
|
||||
/// <summary>
|
||||
/// The Mistral models which carry no further family name: Mistral 7B, Mistral 3, and their kin.
|
||||
/// </summary>
|
||||
/// <remarks>
|
||||
/// The open weights are where these live. Mistral's own API sells the named ranges -- Small,
|
||||
/// Medium, Large -- while the plain checkpoints are the ones people run themselves, which is why
|
||||
/// nothing here is dated: those names carry a size and a quantization instead of a release.
|
||||
///
|
||||
/// A substring, and it has to be one: this is the fallback of the whole range, and every family
|
||||
/// with a name of its own beats it by saying more. What it must not do is claim Ministral or
|
||||
/// Magistral, and it does not -- neither of those two names contains "mistral".
|
||||
/// </remarks>
|
||||
public sealed class MistralFamily : ModelFamily
|
||||
{
|
||||
/// <inheritdoc />
|
||||
public override ModelVendor Vendor => ModelVendor.MISTRAL_AI;
|
||||
|
||||
/// <inheritdoc />
|
||||
public override ModelSource Source => new("https://docs.mistral.ai/getting-started/models/weights/", new DateOnly(2026, 9, 11), "Ported unchanged from the Mistral block of ProviderExtensions.OpenSource.cs: its default answer, and the rule for the 3 line.");
|
||||
|
||||
/// <inheritdoc />
|
||||
protected override void Declare(ModelFamilyBuilder builder)
|
||||
{
|
||||
builder.Rule("mistral").AsSubstring()
|
||||
.Capabilities(TEXT_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
|
||||
.Apis(CHAT_COMPLETION_API);
|
||||
|
||||
// The 3 line reads images and thinks when it is asked to:
|
||||
builder.Rule("mistral-3").AsSegment().Inherits()
|
||||
.Capabilities(MULTIPLE_IMAGE_INPUT)
|
||||
.Reasoning(ReasoningSupport.OPTIONAL);
|
||||
}
|
||||
}
|
||||
@@ -49,7 +49,9 @@ public static class MistralReleases
|
||||
/// <remarks>
|
||||
/// Mistral serves some models under their marketing version as well, and writes the version
|
||||
/// separator both ways: mistral-medium-3.5 and mistral-medium-3-5 are the same model. Those
|
||||
/// names carry no release date, so they are mapped onto the release they stand for.
|
||||
/// names carry no release date, so they are mapped onto the release they stand for. Ollama
|
||||
/// leaves the separator out altogether for the Small checkpoints, which is a third spelling of
|
||||
/// the same statement.
|
||||
///
|
||||
/// The order matters, and it is the one place in this rebuild where it still does: these are
|
||||
/// read as plain text rather than as patterns, so "mistral-medium-3" would answer for
|
||||
@@ -71,6 +73,11 @@ public static class MistralReleases
|
||||
("mistral-small-3.1", 2503),
|
||||
("mistral-small-3-1", 2503),
|
||||
("mistral-small-3", 2501),
|
||||
|
||||
("mistral-small4", 2603),
|
||||
("mistral-small3.2", 2506),
|
||||
("mistral-small3.1", 2503),
|
||||
("mistral-small3", 2501),
|
||||
];
|
||||
|
||||
/// <summary>
|
||||
|
||||
@@ -5,6 +5,12 @@ namespace AIStudio.Models.Mistral;
|
||||
/// <summary>
|
||||
/// Mistral Small, which gained images with 3.1 and reasoning with 4.
|
||||
/// </summary>
|
||||
/// <remarks>
|
||||
/// The one family of the range whose name arrives glued to its version: Ollama publishes the open
|
||||
/// weights as "mistral-small3.1" and "mistral-small3.2", without the separator Mistral's own API
|
||||
/// writes. A substring covers both spellings, and it stays specific enough that nothing else in the
|
||||
/// range can be mistaken for it.
|
||||
/// </remarks>
|
||||
public sealed class MistralSmallFamily : MistralReleaseDatedFamily
|
||||
{
|
||||
/// <inheritdoc />
|
||||
@@ -21,7 +27,7 @@ public sealed class MistralSmallFamily : MistralReleaseDatedFamily
|
||||
|
||||
/// <inheritdoc />
|
||||
protected override void Declare(ModelFamilyBuilder builder) =>
|
||||
builder.Rule("mistral-small").AsSegment()
|
||||
builder.Rule("mistral-small").AsSubstring()
|
||||
.Capabilities(WHAT_THEY_COULD_ALWAYS_DO)
|
||||
.Apis(CHAT_COMPLETION_API);
|
||||
}
|
||||
@@ -0,0 +1,82 @@
|
||||
using AIStudio.Models.Matching;
|
||||
|
||||
using static AIStudio.Provider.Capability;
|
||||
|
||||
namespace AIStudio.Models.ZAI;
|
||||
|
||||
/// <summary>
|
||||
/// GLM, from Z AI.
|
||||
/// </summary>
|
||||
/// <remarks>
|
||||
/// Two things about these names need saying. Z AI writes the version with a dot, but Mistral serves
|
||||
/// the same models as "glm-5-2" and "zai-glm-5-2", so each generation is stated in both spellings.
|
||||
/// And a vision model is marked by a "v" glued to the version number -- glm-4v, glm-4.1v, glm-4.5v
|
||||
/// -- which is not a name part and therefore not something a pattern can ask about. That is what
|
||||
/// the refinement below is for.
|
||||
///
|
||||
/// Looking for a bare "v" anywhere, which the previous rules started out doing, calls every
|
||||
/// quantized build a vision model: "nvfp4" carries one, and so does the name of more than one
|
||||
/// inference provider. The digit in front is what makes it a version marker.
|
||||
/// </remarks>
|
||||
public sealed class GlmFamily : ModelFamily
|
||||
{
|
||||
/// <inheritdoc />
|
||||
public override ModelVendor Vendor => ModelVendor.Z_AI;
|
||||
|
||||
/// <inheritdoc />
|
||||
public override ModelSource Source => new("https://huggingface.co/zai-org", new DateOnly(2026, 9, 11), "Ported unchanged from the Z AI block of ProviderExtensions.OpenSource.cs.");
|
||||
|
||||
/// <inheritdoc />
|
||||
public override ModelProfile Refine(in ModelId id, in ModelProfile selected)
|
||||
{
|
||||
if (!MarksAVisionModel(id.Normalized.AsSpan()))
|
||||
return selected;
|
||||
|
||||
return selected with { Capabilities = selected.Capabilities | MULTIPLE_IMAGE_INPUT };
|
||||
}
|
||||
|
||||
/// <inheritdoc />
|
||||
protected override void Declare(ModelFamilyBuilder builder)
|
||||
{
|
||||
// Every other GLM thinks when the request asks it to:
|
||||
builder.Rule("glm").AsSubstring()
|
||||
.Capabilities(TEXT_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
|
||||
.Apis(CHAT_COMPLETION_API)
|
||||
.Reasoning(ReasoningSupport.OPTIONAL);
|
||||
|
||||
// The 4 line answers straight away:
|
||||
builder.Rule("glm-4").AsSegment()
|
||||
.Capabilities(TEXT_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
|
||||
.Apis(CHAT_COMPLETION_API);
|
||||
|
||||
// 5.2 thinks unless it is told not to:
|
||||
builder.Rule("glm-5.2").AsSegment()
|
||||
.Capabilities(TEXT_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
|
||||
.Apis(CHAT_COMPLETION_API)
|
||||
.Reasoning(ReasoningSupport.ON_BY_DEFAULT);
|
||||
|
||||
builder.Rule("glm-5-2").AsSegment().Inherits();
|
||||
|
||||
// 5.3 thinks whatever it is told: only the effort can be lowered, not the thinking itself.
|
||||
builder.Rule("glm-5.3").AsSegment()
|
||||
.Capabilities(TEXT_INPUT | MULTIPLE_IMAGE_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
|
||||
.Apis(CHAT_COMPLETION_API)
|
||||
.Reasoning(ReasoningSupport.ALWAYS);
|
||||
|
||||
builder.Rule("glm-5-3").AsSegment().Inherits();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Whether the version number of this name is followed by the vision marker.
|
||||
/// </summary>
|
||||
/// <param name="modelName">The normalized model name.</param>
|
||||
/// <returns>True, when a "v" sits directly behind a digit.</returns>
|
||||
private static bool MarksAVisionModel(ReadOnlySpan<char> modelName)
|
||||
{
|
||||
for (var index = 1; index < modelName.Length; index++)
|
||||
if (modelName[index] is 'v' && char.IsAsciiDigit(modelName[index - 1]))
|
||||
return true;
|
||||
|
||||
return false;
|
||||
}
|
||||
}
|
||||
Reference in new issue
Block a user