Say at the model where Codestral cannot be chatted with

This commit is contained in:
Thorsten Sommer 2026-09-19 11:52:40 +02:00
parent adcf118108
commit 1c9ec4d99d
Signed by untrusted user who does not match committer: tsommer
GPG Key ID: 371BBA77A02C0108
3 changed files with 38 additions and 4 deletions

View File

@ -1,3 +1,5 @@
using AIStudio.Provider;
using static AIStudio.Provider.Capability;
namespace AIStudio.Models.Mistral;
@ -10,6 +12,12 @@ namespace AIStudio.Models.Mistral;
/// words the Mistral block looked for, so it walked past every rule and reached the answer meant
/// for everything nobody had written one for. That answer happened to describe it correctly, which
/// is why nothing looked wrong -- and is exactly the situation this rebuild is meant to end.
///
/// That Mistral serves it to fill in the middle of a file rather than to talk to was the one thing
/// the rebuild left behind: it stayed in the Mistral provider, as a name check which dropped every
/// model whose ID begins with "code". It says something about a model, so it belongs to the model,
/// and it is bound to the provider because it is only true there. Somebody's own server and the
/// gateways serve the same weights to chat with, which is what the unbound rule above keeps saying.
/// </remarks>
public sealed class CodestralFamily : ModelFamily
{
@ -17,11 +25,22 @@ public sealed class CodestralFamily : ModelFamily
public override ModelVendor Vendor => ModelVendor.MISTRAL_AI;
/// <inheritdoc />
public override ModelSource Source => new("https://docs.mistral.ai/getting-started/models/models_overview/", new DateOnly(2026, 9, 11), "The answer the previous rules gave it through their fallback: text in, text out, tool calling.");
public override ModelSource Source => new("https://docs.mistral.ai/getting-started/models/models_overview/", new DateOnly(2026, 9, 19), "The answer the previous rules gave it through their fallback: text in, text out, tool calling. What Mistral's own catalog makes of it was read from the same page, where Codestral is the model behind the FIM endpoint.");
/// <inheritdoc />
protected override void Declare(ModelFamilyBuilder builder) =>
protected override void Declare(ModelFamilyBuilder builder)
{
builder.Rule("codestral").AsSegment()
.Capabilities(TEXT_INPUT | TEXT_OUTPUT | FUNCTION_CALLING)
.Apis(CHAT_COMPLETION_API);
//
// Bound to Mistral, which makes it the more specific of the two and lets it win there
// without anybody writing an order. It continues a text instead of answering in a
// conversation, so it must not stand among the models somebody picks for a chat.
//
builder.Rule("codestral").AsSegment().OnlyOn(LLMProviders.MISTRAL)
.Inherits()
.Kind(ModelKind.TEXT_COMPLETION);
}
}

View File

@ -115,7 +115,7 @@ LITE_LLM | anthropic/claude-sonnet-5 | CHAT_COMPLETION_API, FUNCTION_CALLING, MU
LITE_LLM | azure/gpt-5.6 | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, REASONING_BY_DEFAULT, TEXT_INPUT, TEXT_OUTPUT, WEB_SEARCH | CHAT | 1050000 | (unknown) | TIKTOKEN o200k_base
LITE_LLM | bedrock/anthropic.claude-3-5-sonnet-20241022-v2:0 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
LITE_LLM | the-fast-one | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
MISTRAL | codestral-2508 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
MISTRAL | codestral-2508 | CHAT_COMPLETION_API, FUNCTION_CALLING, TEXT_INPUT, TEXT_OUTPUT | TEXT_COMPLETION | (unknown) | (unknown) | (unknown)
MISTRAL | magistral-medium-2506 | ALWAYS_REASONING, CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
MISTRAL | ministral-14b-2512 | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)
MISTRAL | ministral-3b-latest | CHAT_COMPLETION_API, FUNCTION_CALLING, MULTIPLE_IMAGE_INPUT, TEXT_INPUT, TEXT_OUTPUT | CHAT | (unknown) | (unknown) | (unknown)

View File

@ -33,6 +33,11 @@ public static class ModelKindCorpus
// The one name whose only marker used to be the organization it was published under:
new(SELF_HOSTED, "sentence-transformers/all-MiniLM-L6-v2", EMBEDDING),
// Mistral's embedding checkpoint for code. It carries the name of a family which is a text
// completion model on this very provider, and stays an embedding model regardless: what a
// model is for is said by the word which says it, not by the family it was built from.
new(MISTRAL, "codestral-embed", EMBEDDING),
];
/// <summary>
@ -128,13 +133,18 @@ public static class ModelKindCorpus
];
/// <summary>
/// The models from before chat completions existed.
/// The models which continue a text instead of answering in a conversation.
/// </summary>
private static readonly ModelKindExample[] TEXT_COMPLETION_ENTRIES =
[
new(HELMHOLTZ, "text-davinci-003", TEXT_COMPLETION),
new(OPEN_AI, "babbage-002", TEXT_COMPLETION),
new(OPEN_AI, "gpt-3.5-turbo-instruct", TEXT_COMPLETION),
// The one of these which is not old: Mistral serves Codestral to fill in the middle of a
// file. It says so only for Mistral's own catalog, which is why the open weights of the
// same name stay a chat model further down.
new(MISTRAL, "codestral-latest", TEXT_COMPLETION),
];
/// <summary>
@ -179,6 +189,11 @@ public static class ModelKindCorpus
new(SELF_HOSTED, "llama3.3:70b", CHAT),
new(OPEN_AI, "gpt-5.1", CHAT),
// The open weights of the model Mistral itself serves to fill in the middle of a file.
// Whoever runs them runs them behind a chat completion API, so here the name means
// something to talk to -- which is what binding that other rule to Mistral protects.
new(SELF_HOSTED, "codestral-22b-v0.1", 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