Model

Stable Beluga 1 (FreeWilly 1) 65B

OPEN

Stability AI (CarperAI) · Stable Beluga (FreeWilly) family · released Jul 21, 2023

Ranked #4 on the HF Open LLM Leaderboard at launch (Jul 27 2023); released as delta weights over Meta's LLaMA 65B.

ReasoningCodingVisionFunction callingTool useAgentic

177.8

Elo · rank #369

Parameters

65B

Active params

65B (dense)

Context

2K tokens

Architecture

LLaMA 65B dense Transformer, Orca-style supervised fine-tune

License

CC BY-NC-4.0 (Non-Commercial Creative Commons License) — released as delta weights over Meta's LLaMA 65B

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

ARC-ChallengeReasoning68.2%#56

best: Llama 3.1 405B · 96.9%

DROPReasoning0#79

best: Hunyuan-T1 · 93.1

GSM8KMath0.0%#203

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning85.9%#34

best: Claude 3 Opus · 95.4%

MMLUKnowledge64.8%#180

best: OpenAI o3 · 92.9%

TruthfulQAKnowledge55.8%#36

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WinoGrandeReasoning49.8%#139

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

130 GB

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Stable Beluga (FreeWilly) family

Elo progression across releases

Stable Beluga 1 (FreeWilly 1) 65BJul 2023177.8 Stable Beluga 2 (FreeWilly 2) 70BJul 2023511.0

Resources

HFstabilityai/StableBeluga1-Delta

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