Model
Model explorer
OPT-IML 30B
OPENMeta · OPT family · released Dec 22, 2022
Instruction-tuned OPT fine-tuned on a curated ~1,500-task subset of the OPT-IML Bench, with the remaining tasks held out for downstream evaluation; early instruction-tuning at scale. (OPT-IML-Max is a separate sibling checkpoint trained on the full ~2,000-task bench.)
ReasoningCodingVisionFunction callingTool useAgentic
-189.2
Elo · unrated
Parameters
30B
Active params
30B (dense)
Context
2.048K tokens
Architecture
Dense decoder-only Transformer (OPT architecture), instruction-tuned
License
OPT-175B License Agreement (non-commercial research use)
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
best: Llama 3.1 405B · 96.9%
best: Phi-3-medium (14B) · 97.7%
best: Claude 1 · 90.8%
best: GPT-4o mini · 93.1%
best: GPT-3 175B · 87.7%
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
17 GB
VRAM @ FP16
60 GB
Fits on (Q4)
RTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
llama.cpp does not support the OPT architecture, so no RTX 4090 llama.cpp figure exists even though the Q4 footprint would fit in 24GB.
Quantizations
GPTQ
OPT family
Elo progression across releases
API price weights · each benchmark row carries its own source badge (see methodology)