Skip to content
Model infoModelsSnapshot: 15 Oct 2026

Apple / Apple Foundation Models 2025

AFM On-Device (2025)

Second-generation multimodal Apple foundation model. Model-level image capability is distinct from what the original developer framework exposed.

Language models (LLMs)
historicalClosed weightsText generation
Official documentation (opens in a new tab)Reviewed 14 Oct 2026
16 of 21 specification fields collected
Context window4.1K4,096 tokens
Maximum outputNot collectedAwaiting source review
Input billingSee billing notesApple Foundation Models framework (on-device)
Output billingSee billing notesAccess and billing notes below

Specifications

Provider-documented facts. “Not collected” means we have not verified that field yet, not that the feature is unsupported.

Context / input limit
4,096 tokensSource for Context / input limit (opens in a new tab)

Shared session budget for instructions, prompts, tool definitions/results, schemas and generated responses. Query contextSize at runtime; this is not a separate 4,096-token output allowance.

Maximum input tokens
Not collected
Maximum output
Not collected
Input modalities
Text, ImageSource for Input modalities (opens in a new tab)
Output modalities
TextSource for Output modalities (opens in a new tab)
Reasoning
Not collected
Tool calling
SupportedSource for Tool calling (opens in a new tab)

Framework tool calling, with developer-provided Tool implementations.

Structured output
SupportedSource for Structured output (opens in a new tab)

Guided generation using Swift Generable schemas and constrained decoding.

Streaming
SupportedSource for Streaming (opens in a new tab)

Response snapshots, including partially generated structured objects.

Batch processing
Not collected
Fine-tuning
SupportedSource for Fine-tuning (opens in a new tab)

Version-matched LoRA adapters, not full base-weight training. Toolkit 26.0.0 is the final release; incompatible with OS 27 and later. Training downloads require Developer Program membership and deployment requires an entitlement.

Knowledge cutoff
Not collected
Release date
2025-06-09Source for Release date (opens in a new tab)

Generation announcement.

Parameters
Approximately 3B language modelSource for Parameters (opens in a new tab)
Architecture
Two decoder blocks in a 5:3 depth ratio, sharing KV state across blocksSource for Architecture (opens in a new tab)
Languages
16 languages in technical report; overview says 15Source for Languages (opens in a new tab)

Preserve the source discrepancy. Supported-language count is not a promise that every feature supports every locale.

Context caching
SupportedSource for Context caching (opens in a new tab)

Session-local KV reuse and optional prompt prewarming. This does not imply a shared server prompt cache or a cached-token price.

Quantization and compression
2-bit decoder QAT; 4-bit embeddings; 8-bit KV cacheSource for Quantization and compression (opens in a new tab)
Vision encoder
300M-parameter ViTDet-LSource for Vision encoder (opens in a new tab)

Component size, not the language decoder or total model size.

Published price
No per-token cloud inference charge for on-device useSource for Published price (opens in a new tab)

Requires an eligible Apple device and enabled/downloaded Apple Intelligence model. Hardware, distribution and developer membership costs are separate.

Customization options
Rank-32 LoRA adapters; optional draft-model trainingSource for Customization options (opens in a new tab)

Access & pricing

Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.

Apple Foundation Models framework (on-device)

Provider source (opens in a new tab)
API model ID
Not collected

Product or SDK model identity, not a verified public HTTP model selector.

Access
Swift SystemLanguageModel.default selects the installed OS model. There is no verified user-selectable checkpoint pin or public hosted REST model ID.
Regions
Supported Apple devices, OS versions, languages and regions only. A device-local model is not a guarantee of access in every market.
SDK selector
SystemLanguageModel.defaultSource for SDK selector (opens in a new tab)

SDK type/member, not an HTTP model ID.

Installed model version
OS 26.0–26.3Source for Installed model version (opens in a new tab)
Input modalities
TextSource for Input modalities (opens in a new tab)

Original developer framework exposed text tasks, even though the research model includes image understanding.

Endpoint pricing has not been collected. Check the provider for applicable rates.

Data & enterprise terms

Retention, training use, regional processing and enterprise terms have not been summarized for this endpoint.

Caching

Caching can reuse earlier input processing. Availability, lifetime and billing depend on the endpoint. A cache hit still generates a new response.

01 / INITIAL INPUTProcess & retain

The system processes eligible input and retains reusable state.

02 / LATER REQUESTCheck available state

Reuse depends on matching input, session state and the cache policy.

03 / CACHE HITReuse & generate

Reusing state can reduce repeated work. Any charges follow the billing rules below.

Apple Foundation Models framework (on-device)

Documented

Session-local KV reuse and optional prompt prewarming. This does not imply a shared server prompt cache or a cached-token price.

Support evidence (opens in a new tab)
How it works
Stateful session KV cache; optional eager prompt-prefix processingSource for How it works (opens in a new tab)
Minimum reusable input
Not collected

No numeric minimum reusable prefix is documented here.

Lifetime & refresh
Not collected

No fixed TTL or cross-session persistence guarantee verified.

How to enable it
Call LanguageModelSession.prewarm(promptPrefix:) before an expected requestSource for How to enable it (opens in a new tab)
Usage fields
Allow at least one second before responding; eager loading is not guaranteed under load or in backgroundSource for Usage fields (opens in a new tab)
Billing rules
Local inference; no published cached-token tariffSource for Billing rules (opens in a new tab)
Limits & data boundaries
Session state and hardware resources govern reuse; not evidence of paid API cache behaviorSource for Limits & data boundaries (opens in a new tab)

No verified cache rate for this endpoint. An empty rate is not a zero charge.

Caching reviewed 14 Oct 2026

What the measurements say

MMLU

Apple · vendor-reported
67.85%

Apple technical report Tables 1–2, instruct/non-thinking models and Simple-Evals on supported languages. These are paper model results, not guaranteed scores for the compressed installed OS build.

Evaluation source (opens in a new tab)

MMMLU

Apple · vendor-reported
60.6%

Apple technical report Tables 1–2, instruct/non-thinking models and Simple-Evals on supported languages. These are paper model results, not guaranteed scores for the compressed installed OS build.

Evaluation source (opens in a new tab)

MGSM

Apple · vendor-reported
74.91%

Apple technical report Tables 1–2, instruct/non-thinking models and Simple-Evals on supported languages. These are paper model results, not guaranteed scores for the compressed installed OS build.

Evaluation source (opens in a new tab)

Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.

The record behind the model

Why it is included

Apple documents this proprietary product model. No matching unrestricted public checkpoint was verified in the reviewed Apple weight inventory. Closed-weight classification is inferred for this product identity; OS delivery and member-gated model assets do not imply an open-weight release.

Classification evidence (opens in a new tab)

Lifecycle dates

Generation announced9 Jun 2025Source (opens in a new tab)

Sources reviewed