Support Apple Intelligence as an on-device AI model for the voice assistant/etc

This commit is contained in:
James Babcock committed 2026-10-31 17:48:45 -07:00
1 parent 3c3aa31542
commit 4973fd2aea
22 files changed
+1551 -40

No files matched your search

+2 -2
View File
@@ -1,6 +1,6 @@
import { streamAnthropicMessage } from "../native/anthropic";
import { streamHermesChat } from "../native/hermes";
import { streamLocalQwen } from "../native/llama";
import { streamLocalModel } from "../native/local-llm";
import { streamOpenAiResponse } from "../native/openai";
import type {
LlmContentBlock,
@@ -146,7 +146,7 @@ export class DirectAssistantBackend {
: options.provider === "openai"
? streamOpenAiResponse(streamOptions)
: options.provider === "local"
? streamLocalQwen(streamOptions)
? streamLocalModel(streamOptions)
: streamAnthropicMessage(streamOptions);
};
+13 -7
View File
@@ -1,4 +1,4 @@
import { isLocalModelReady, LOCAL_MODEL } from "../native/llama";
import { isLocalModelReady, isLocalModelSupported, LOCAL_MODEL } from "../native/local-llm";
export const ASSISTANT_MODEL_VALUES = [
"auto",
@@ -13,12 +13,17 @@ export const ASSISTANT_MODEL_VALUES = [
"hermes",
] as const;
/**
* "qwen" is the on-phone model, whatever it is on this platform: Qwen3-4B on
* Android, the Apple Intelligence system model on iOS (the ID predates iOS
* support and stays for saved settings and conversations).
*/
export type AssistantModel = (typeof ASSISTANT_MODEL_VALUES)[number];
/** Keep saved model IDs portable, while offering only supported choices. */
export const ASSISTANT_MODEL_CHOICES: readonly AssistantModel[] =
global.isIOS ? ASSISTANT_MODEL_VALUES.filter(value => value !== "qwen") : ASSISTANT_MODEL_VALUES;
isLocalModelSupported() ? ASSISTANT_MODEL_VALUES : ASSISTANT_MODEL_VALUES.filter(value => value !== "qwen");
export function supportedAssistantModel(value: AssistantModel): AssistantModel {
return global.isIOS && value === "qwen" ? "auto" : value;
return value === "qwen" && !isLocalModelSupported() ? "auto" : value;
}
export type AssistantProvider = "anthropic" | "openai" | "local" | "hermes";
@@ -58,8 +63,9 @@ const MODEL_DEFINITIONS: Record<Exclude<AssistantModel, "auto">, ModelDefinition
luna: { label: "Luna", provider: "openai", model: "gpt-5.6-luna", effort: "low" },
terra: { label: "Terra", provider: "openai", model: "gpt-5.6-terra", effort: "low" },
sol: { label: "Sol", provider: "openai", model: "gpt-5.6-sol", effort: "low" },
// Runs on the phone itself (llama.cpp); needs the downloaded model file,
// not an API key. The default for users who haven't set any key.
// Runs on the phone itself (Android: llama.cpp with the downloaded model
// file; iOS: Apple Intelligence), not an API key. The default for users who
// haven't set any key.
qwen: { label: `${LOCAL_MODEL.label} (on-phone)`, provider: "local", model: LOCAL_MODEL.id },
// Your own Hermes Agent (its API server), with its own tools and memory.
hermes: { label: "Hermes agent", provider: "hermes", model: "hermes-agent" },
@@ -76,8 +82,8 @@ export function assistantModelProvider(value: AssistantModel): AssistantProvider
/**
* Resolve Auto and attach the credential needed by the chosen provider.
* Auto intentionally prefers Terra over Sonnet when both keys are available,
* and falls back to the on-phone model when no key is set but the model file
* has been downloaded.
* and falls back to the on-phone model when no key is set but that model is
* ready (downloaded on Android; Apple Intelligence on and ready on iOS).
*/
export function resolveAssistantModel(
selection: AssistantModel,