Pythia models
15 Pythia models from EleutherAI on Hugging Face, from 213M to 12.0B parameters, published by EleutherAI in F16 and F32, with a 2K-token context. The smallest official model, pythia-160m, needs about 0.5 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Pick a size
One row per official Pythia size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| pythia-160m | 213M | 0.5 GB | 0.2 GB | 0.1 GB | RTX 3070 | $0.088/hr | 1.1 GB |
| pythia-410m | 506M | 1.1 GB | 0.6 GB | 0.3 GB | RTX 3070 | $0.088/hr | 3.0 GB |
| pythia-1b | 1.1B | 2.4 GB | 1.2 GB | 0.6 GB | RTX 3070 | $0.088/hr | 4.0 GB |
| pythia-1.4b-deduped | 1.4B | 6.3 GB | 1.6 GB | 0.8 GB | RTX 3070 | $0.088/hr | 6.0 GB |
| pythia-1.4b | 1.5B | 3.4 GB | 1.7 GB | 0.8 GB | RTX 3070 | $0.088/hr | 6.0 GB |
| pythia-2.8b | 2.9B | 6.5 GB | 3.3 GB | 1.6 GB | RTX 3070 | $0.088/hr | 10.0 GB |
| pythia-6.9b | 7.0B | 15.6 GB | 7.8 GB | 3.9 GB | V100 | $0.110/hr | 16.0 GB |
| pythia-12b | 12.0B | 26.8 GB | 13.4 GB | 6.7 GB | V100 | $0.220/hr | 22.5 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (14)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| pythia-160m | 213M | F16 | 0.5 GB | RTX 3070 | $0.088/hr |
| pythia-160m-deduped | 213M | F16 | 0.5 GB | RTX 3070 | $0.088/hr |
| pythia-160m-seed1 | 213M | F16 | 0.5 GB | RTX 3070 | $0.088/hr |
| pythia-160m-seed2 | 213M | F16 | 0.5 GB | RTX 3070 | $0.088/hr |
| pythia-160m-seed3 | 213M | F16 | 0.5 GB | RTX 3070 | $0.088/hr |
| pythia-410m | 506M | F16 | 1.1 GB | RTX 3070 | $0.088/hr |
| pythia-410m-deduped | 506M | F16 | 1.1 GB | RTX 3070 | $0.088/hr |
| pythia-1b | 1.1B | F16 | 2.4 GB | RTX 3070 | $0.088/hr |
| pythia-1b-deduped | 1.1B | F16 | 2.4 GB | RTX 3070 | $0.088/hr |
| pythia-1.4b-deduped | 1.4B | F32 | 6.3 GB | RTX 3070 | $0.088/hr |
| pythia-1.4b | 1.5B | F16 | 3.4 GB | RTX 3070 | $0.088/hr |
| pythia-2.8b | 2.9B | F16 | 6.5 GB | RTX 3070 | $0.088/hr |
| pythia-6.9b | 7.0B | F16 | 15.6 GB | V100 | $0.110/hr |
| pythia-12b | 12.0B | F16 | 26.8 GB | V100 | $0.220/hr |
Fine-tunes and community models (1)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| pythia410m-sft-tldr | 405M | F32 | 1.8 GB | RTX 3070 | $0.088/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.