Model
Model explorer

LLaMA 33B

OPEN
Meta · LLaMA family · released Feb 24, 2023

Mid-large LLaMA-1 checkpoint (32.5B params) trained on 1.4T tokens, sitting between the 13B and 65B sizes.

ReasoningCodingVisionFunction callingTool useAgentic
184.9
Elo · rank #366
Parameters
32.5B
Active params
32.5B (dense)
Context
2.048K tokens
Architecture
Dense decoder-only Transformer (RMSNorm, SwiGLU, rotary position embeddings)
License
LLaMA License (non-commercial research, gated/case-by-case access)
Languages
20+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning57.8%#99
best: Llama 3.1 405B · 96.9%
ARC-EasyReasoning80.0%#26
best: Phi-3-medium (14B) · 97.7%
FIN-benchKnowledge53.4%#2
best: Poro 34B · 66.3%
GSM8KMath35.6%#147
best: Llama 3.1 405B · 96.8%
HellaSwagReasoning82.8%#60
best: Claude 3 Opus · 95.4%
HumanEvalCoding21.7%#167
best: Claude Opus 4.5 · 99.4%
MBPPCoding30.2%#98
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
MMLUKnowledge57.8%#213
best: OpenAI o3 · 92.9%
OpenBookQAReasoning58.6%#15
best: Claude 1 · 90.8%
PIQAReasoning82.3%#26
best: GPT-4o mini · 93.1%
RACE-HReasoning48.3%#13
best: Claude 3 Opus · 92.9%
Social IQaReasoning50.4%#26
best: Apple DCLM-Baseline 7B · 82.9%
TriviaQAKnowledge65.1%#28
best: Sarvam-1 (2B) · 90.6%
WinoGrandeReasoning76.0%#59
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
19 GB
VRAM @ FP16
65 GB
Fits on (Q4)
RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Original LLaMA-1 predates most public RTX 4090 llama.cpp benchmarks (which target Llama 2/3); no verified figure for this exact checkpoint was found.
Quantizations
GGUF Q4_K_M · GGML · GPTQ · AWQ
LLaMA family
Elo progression across releases
LLaMA 13BFeb 20230.6LLaMA 33BFeb 2023184.9LLaMA 65BFeb 2023240.9LLaMA 7BFeb 2023-180.9
Resources
HFhuggyllama/llama-33bGHfacebookresearch/llamaDOCSarxiv.org/abs/2302.13971
API price weights · each benchmark row carries its own source badge (see methodology)