Model

Apple DCLM-Baseline 7B

OPEN

Apple · Apple DCLM family · released Jul 1, 2024

Data-centric base LM trained on the curated DCLM-Baseline corpus (3.8T tokens + StarCoder/ProofPile2); not instruction-tuned.

ReasoningCodingVisionFunction callingTool useAgentic

62.7

Elo · rank #397

Parameters

Active params

7B (dense)

Context

2K tokens

Architecture

Dense decoder-only Transformer (32 layers, 4096 hidden, 32 heads)

License

Apple Sample Code License

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning59.9%#91

best: Llama 3.1 405B · 96.9%

ARC-EasyReasoning82.2%#18

best: Phi-3-medium (14B) · 97.7%

GPQA DiamondReasoning24.8%#286

best: GPT-6 Astra · 96.0%

GSM8KMath2.5%#190

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning80.4%#77

best: Claude 3 Opus · 95.4%

LAMBADAReasoning82.2%#2

best: GPT-3 175B · 86.4%

MMLUKnowledge63.7%#184

best: OpenAI o3 · 92.9%

OpenBookQAReasoning45.4%#36

best: Claude 1 · 90.8%

PIQAReasoning81.3%#32

best: GPT-4o mini · 93.1%

PubMedQAKnowledge70.6%#8

best: Palmyra Med 70B · 79.6%

Social IQaReasoning82.9%#1

best: this model · 82.9%

TriviaQAKnowledge52.7%#42

best: Sarvam-1 (2B) · 90.6%

WinoGrandeReasoning72.7%#84

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

14 GB

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Conditional / custom

QLoRA6.4 GB1× RTX 3060 12GB

LoRA16.6 GB1× RTX 3090 24GB

Full fine-tune114.0 GB1× H200 141GB

QLoRA SFT on ~10k samples ≈ $4.10 (1× RTX 3060 12GB)

Plan a fine-tune →

Apple DCLM family

Elo progression across releases

Apple DCLM-Baseline 7BJul 202462.7

Resources

HFapple/DCLM-Baseline-7B

API price weights · each benchmark row carries its own source badge (see methodology)