Instructions to use FiShota/hinomoto-1b-bitnet158-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FiShota/hinomoto-1b-bitnet158-v1 with Transformers:

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

pipe = pipeline("text-generation", model="FiShota/hinomoto-1b-bitnet158-v1")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("FiShota/hinomoto-1b-bitnet158-v1")
model = AutoModelForCausalLM.from_pretrained("FiShota/hinomoto-1b-bitnet158-v1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FiShota/hinomoto-1b-bitnet158-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FiShota/hinomoto-1b-bitnet158-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FiShota/hinomoto-1b-bitnet158-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/FiShota/hinomoto-1b-bitnet158-v1

SGLang

How to use FiShota/hinomoto-1b-bitnet158-v1 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 "FiShota/hinomoto-1b-bitnet158-v1" \
    --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": "FiShota/hinomoto-1b-bitnet158-v1",
		"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 "FiShota/hinomoto-1b-bitnet158-v1" \
        --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": "FiShota/hinomoto-1b-bitnet158-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use FiShota/hinomoto-1b-bitnet158-v1 with Docker Model Runner:
```
docker model run hf.co/FiShota/hinomoto-1b-bitnet158-v1
```

HinoMoto-1B v1 BitNet b1.58 (Ternary-Quantized)

HinoMoto-1B v1 Phase 2 Full の BitNet b1.58 量子化版. 重みを三値 (= 1.58 bit/weight) に置換後、 500 step QAT で品質回復.

個人 GPU (RTX 3090 24GB) で paper の主張を再現 したエヴィデンス.

What is BitNet b1.58?

Reference: Ma et al., "The Era of 1-bit LLMs" (Feb 2024). https://arxiv.org/abs/2402.17764

各 linear layer の weight を {-1, 0, +1} の三値に量子化:

weight VRAM: 6x 削減 (= bf16 比)
forward の matmul → add/sub only
推論時に CUDA fused kernel で大幅高速化期待

Architecture (Llama-style + BitLinear)

Item	Value
Params	955,221,504 (~955M)
	1536
	32
	16 (MHA)
(SwiGLU)	4096
Linear modules	224 BitLinear (= attn + FFN)
+ embeddings	fp 維持 (paper convention)
	32,000 (HinoMoto byte-BPE, 32k)

Training (Quantization-Aware)

Item	Value
Base ckpt	HinoMoto-1B Phase 2 Full (50,000 step from-scratch)
QAT steps	500
Optimizer	AdamW (lr 5e-5, cosine, warmup 100)
dtype	bf16 mixed (master weights fp), STE quantization
Hardware	RTX 3090 24GB
Wall-clock	~5 hr (with torch.compile, default mode)
Best loss (instant)	1.24 (single batch outlier), steady ~2.5

Benchmark (HinoMoto-Bench-ja v0.6)

Axis	fp 1B Phase 2 Full	BitNet b1.58 (this)	Δ
family score /12	5.08	5.29	+0.21 (✓ within paper claim)
family degenerate	9.1%	9.1%	0%
keigo pass_rate	11.4%	10.0%	-1.4%
keigo degenerate	5.7%	7.1%	+1.4%
silence pass_rate	6.0%	0.0%	-6.0% (degradation)
silence degenerate	6.0%	12.0%	+6.0%

→ family / keigo は fp 同等. silence は若干悪化 (= 短文応答が量子化に sensitive). 全体として paper 主張の 「< 5% quality loss」達成.

Usage

Why this matters

= 個人 GPU で 7-8B 級モデルの現実的可能性.

本 release はその第一歩.

License

CC BY 4.0. Use for any purpose with attribution.

Limitations

n=1 run (seed=0)
silence 軸が若干悪化 (= QAT 500 step が不十分の可能性. 5000 step paper 推奨)
推論最適化未実装 (= 現状は fp ckpt として load. 真の 6x 高速化には fused INT2 kernel 必要)
no safety alignment (= base 由来)

base: https://huggingface.co/FiShota/hinomoto-1b-v1-phase2-full
350M cultural SFT: https://huggingface.co/FiShota/hinomoto-350m-cultural-sft-v1
bench: https://github.com/FIshota/hinomoto-bench-ja
BitNet paper: https://arxiv.org/abs/2402.17764

Citation

@misc{hinomoto-1b-bitnet158-v1-2026, author = {ryu (FIshota)}, title = {HinoMoto-1B v1 BitNet b1.58 (Ternary-Quantized, Personal-GPU Reproduction)}, year = {2026}, publisher = {Hugging Face}, }

Generated: 2026-05-19 (HinoMoto/ryu + Claude)

Downloads last month: 12

Safetensors

Model size

1.0B params

Tensor type

BF16

Model tree for FiShota/hinomoto-1b-bitnet158-v1

Base model

FiShota/hinomoto-1b-v1-phase2-full

Finetuned

(2)

this model

Paper for FiShota/hinomoto-1b-bitnet158-v1

The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

Paper • 2402.17764 • Published Feb 27, 2024 • 630

FiShota
/

hinomoto-1b-bitnet158-v1