Instructions to use Ayaka157/RetNet-0.5B-v21.2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ayaka157/RetNet-0.5B-v21.2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Ayaka157/RetNet-0.5B-v21.2")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Ayaka157/RetNet-0.5B-v21.2", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Ayaka157/RetNet-0.5B-v21.2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ayaka157/RetNet-0.5B-v21.2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ayaka157/RetNet-0.5B-v21.2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Ayaka157/RetNet-0.5B-v21.2

SGLang

How to use Ayaka157/RetNet-0.5B-v21.2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Ayaka157/RetNet-0.5B-v21.2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ayaka157/RetNet-0.5B-v21.2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Ayaka157/RetNet-0.5B-v21.2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ayaka157/RetNet-0.5B-v21.2",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Ayaka157/RetNet-0.5B-v21.2 with Docker Model Runner:
```
docker model run hf.co/Ayaka157/RetNet-0.5B-v21.2
```

RetNet-0.5B v21.2 — 16k long-context export (measured negative on retrieval)

Status: research export — 16k language modelling works, 16k retrieval does not.

Repository (code, analysis, report): https://github.com/Ayaka157/RetNet
Technical report (2026-10-29): https://github.com/Ayaka157/RetNet/blob/main/docs/latex/en/RetNet-0.5B-Technical-Report.pdf

Model


Parameters	504,179,896
Architecture	RetNet (retention), 20 layers, width 1408, 22 heads, value/FFN dim 2816
Vocabulary / context	20,000 ByteLevel BPE / 2048 training context
Decay	per-head learned $\gamma$; GLA branch implemented but `gla_rank=0`
Precision	stored as fp32 safetensors (1.92 GB); load with orch_dtype=torch.bfloat16 for inference

Files

config.json, model.safetensors, modeling_retnet.py, retention_fast.py, fused_ce.py, tokenizer files. The modeling files come from the published RetNet codebase and are provided so the checkpoint can be loaded with trust_remote_code=True.

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Ayaka157/RetNet-0.5B-v21.2", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Ayaka157/RetNet-0.5B-v21.2", trust_remote_code=True,
                                             torch_dtype="bfloat16")

For serving/prefill use the parallel form, training uses chunkwise, decoding uses recurrent (RETNET_FAST_CHUNKWISE=1 for batched chunkwise). The demo: https://retnet.ayaka157.icu/play/

Evidence and limitations

Long-context training (13,000 steps, 0.852 B tokens) lowered the 16k within-document long-tail from 3.2841 to 2.5942, but the gain is a uniform level shift (all position buckets improve by -0.68 to -0.72 nat), not a position effect. The learned decays collapsed during training: max half-life 1,649 tokens, no head has a half-life >=2,048, max weight at 16k is 0.00102. Needle retrieval at 16k is 0.0978 (chance; null arm 0.1024, item_acc 0/360). Do not advertise or use this export for 16k retrieval; it is evidence for the report.

Ayaka157
/

RetNet-0.5B-v21.2

RetNet-0.5B v21.2 — 16k long-context export (measured negative on retrieval)

Model

Files

Evidence and limitations

Links