Model

Airoboros 70B (L2)

OPEN

Jon Durbin (independent) · Airoboros family · released Aug 5, 2023

Fully synthetic, GPT-4-generated self-instruct fine-tune of Llama 2 70B; a solo-developer pipeline that produced a broadly competitive open 70B chat model.

ReasoningCodingVisionFunction callingTool useAgentic

473.4

Elo · rank #315

Parameters

70B

Active params

70B (dense)

Context

4K tokens

Architecture

Llama 2 (dense decoder-only transformer)

License

Llama 2 Community License Agreement (custom Meta license)

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

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

ARC-ChallengeReasoning68.5%#54

best: Llama 3.1 405B · 96.9%

DROPReasoning40.63#56

best: Hunyuan-T1 · 93.1

GSM8KMath24.7%#161

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning87.9%#14

best: Claude 3 Opus · 95.4%

MMLUKnowledge70.4%#143

best: OpenAI o3 · 92.9%

TruthfulQAKnowledge49.8%#58

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

WinoGrandeReasoning83.5%#16

best: PaLM 2 · 90.9%

Run it locally

VRAM @ Q4

40 GB

VRAM @ FP16

140 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

70B does not fit in a single RTX 4090's 24GB VRAM at Q4 (~40GB needed); no verified single-GPU llama.cpp measurement found.

Quantizations

GGUF Q4 · GPTQ · GGML · EXL2

Airoboros family

Elo progression across releases

Airoboros 70B (L2)Aug 2023473.4

Resources

HFjondurbin/airoboros-l2-70b-gpt4-2.0 GHjondurbin/airoboros

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