{eliza_short_name}

Local-first Eliza-1 agent checkpoint trained with APOLLO on the elizaOS native JSON-format trajectory corpus.

{eliza_short_name} is a full-parameter SFT Eliza-1 checkpoint trained on elizaos/eliza-native-v1-sft. Trained with APOLLO (full fine-tune at SGD-like memory) using the elizaos/eliza-1-training repo. Upstream lineage is recorded in the model metadata and license block.

Model description

field	value
Upstream lineage	recorded in `base_model` metadata
Parameters	{params_billion}B
Architecture	Eliza-1 hybrid local-inference backbone
Training data	`elizaos/eliza-native-v1-sft`
Training pipeline	`elizaos/eliza-1-training`
Optimizer	{optimizer} (rank {optimizer_rank})
Train sequence length	{seq_len}
Native context window	{infer_max_in_plus_out} tokens ({infer_max_in} in + {infer_max_out} out)
License	Apache-2.0 (inherited from base)

Inference

transformers (bf16)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("{repo_id}")
model = AutoModelForCausalLM.from_pretrained(
    "{repo_id}", torch_dtype=torch.bfloat16, device_map="auto",
)

vLLM

Use the in-repo serve script scripts/inference/serve_vllm.py or invoke vLLM directly:

bash

vllm serve {repo_id} --max-model-len {infer_max_in_plus_out}

eliza

Drop the model alias into your eliza config and the runtime will pick it up:

bash

ELIZA_LOCAL_MODEL={repo_id} eliza run

Training

{training_table}

Evaluation

{eval_table}

License

Apache-2.0, inherited from the upstream base checkpoint recorded in base_model. Eliza-side weights and training pipeline released under the same license; see https://github.com/elizaOS for source. Use of the model is additionally governed by the upstream acceptable-use policy inherited via the base checkpoint.

Citation

bibtex

@misc{{eliza1_{eliza_citation_key},
  title  = {{ {eliza_short_name}: an elizaOS local-first agent }},
  author = {{ elizaOS team }},
  year   = {{ 2026 }},
  url    = {{ https://huggingface.co/{repo_id} }},
}}