Model
Model explorer

Guanaco 33B

OPEN
University of Washington (Dettmers et al.) · Guanaco family · released May 23, 2023

QLoRA 4-bit finetune of LLaMA-33B on OASST1 that reached 97.8% of ChatGPT's Vicuna-benchmark quality (second-best in the paper, after the 65B model's 99.3%) after roughly 12 hours on a single 24GB consumer GPU.

ReasoningCodingVisionFunction callingTool useAgentic
280.6
Elo · rank #350
Parameters
33B
Active params
33B (dense)
Context
2K tokens
Architecture
Dense decoder-only transformer (LLaMA-33B base) with 4-bit QLoRA adapter finetune
License
Apache-2.0 (LoRA adapter); base LLaMA-33B weights under Meta's non-commercial research license
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning62.5%#78
best: Llama 3.1 405B · 96.9%
Arena EloHuman preference1065#49
best: Claude Fable 5 · 1505
HellaSwagReasoning84.5%#46
best: Claude 3 Opus · 95.4%
MMLUKnowledge57.0%#215
best: OpenAI o3 · 92.9%
MT-BenchHuman preference6.53#61
best: Hunyuan-Large (A52B) · 9.4
OpenAssistant BenchHuman preference1002#4
best: GPT-4 · 1294
TruthfulQAKnowledge51.2%#51
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
Vicuna BenchHuman preference1009#3
best: GPT-4 · 1176
Run it locally
VRAM @ Q4
21 GB
VRAM @ FP16
66 GB
Fits on (Q4)
RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF Q4 · GPTQ · AWQ
Guanaco family
Elo progression across releases
Guanaco 33BMay 2023280.6Guanaco 65BMay 2023396.3
Resources
HFtimdettmers/guanaco-33bGHartidoro/qloraDOCSarxiv.org/abs/2305.14314
API price weights · each benchmark row carries its own source badge (see methodology)