Instructions to use axteam/Nanbeige4.3-33B-FP8-dynamic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use axteam/Nanbeige4.3-33B-FP8-dynamic with Transformers:

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

pipe = pipeline("text-generation", model="axteam/Nanbeige4.3-33B-FP8-dynamic")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

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

tokenizer = AutoTokenizer.from_pretrained("axteam/Nanbeige4.3-33B-FP8-dynamic")
model = AutoModelForCausalLM.from_pretrained("axteam/Nanbeige4.3-33B-FP8-dynamic", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use axteam/Nanbeige4.3-33B-FP8-dynamic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "axteam/Nanbeige4.3-33B-FP8-dynamic"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "axteam/Nanbeige4.3-33B-FP8-dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/axteam/Nanbeige4.3-33B-FP8-dynamic

SGLang

How to use axteam/Nanbeige4.3-33B-FP8-dynamic 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 "axteam/Nanbeige4.3-33B-FP8-dynamic" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "axteam/Nanbeige4.3-33B-FP8-dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "axteam/Nanbeige4.3-33B-FP8-dynamic" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "axteam/Nanbeige4.3-33B-FP8-dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use axteam/Nanbeige4.3-33B-FP8-dynamic with Docker Model Runner:
```
docker model run hf.co/axteam/Nanbeige4.3-33B-FP8-dynamic
```

Nanbeige4.3-33B · FP8 (dynamic)

8-bit quantization of Nanbeige/Nanbeige4.3-33B for vLLM serving. Quantized by axteam.


Base model	`Nanbeige/Nanbeige4.3-33B`
Method	FP8 (dynamic)
Format	compressed-tensors `float-quantized` (scheme `FP8_DYNAMIC`)
Weights / activations	W8A8 — FP8 E4M3 weights (per-channel), dynamic per-token FP8 activations
Architecture	Dense (Llama architecture) · 52 layers · text-only
Size on disk	~35.5 GB
Calibration	None (data-free)
License	apache-2.0

Quantization details

Scheme — W8A8 — FP8 E4M3 weights (per-channel), dynamic per-token FP8 activations, stored as compressed-tensors float-quantized (scheme FP8_DYNAMIC).
Kept in BF16 — token embeddings, the LM head, and all RMSNorms. Only attention & MLP projections are quantized.
Post-training quantization (no retraining). Tokenizer / chat_template identical to the original.

Benchmarks (measured)

Every column was measured with the same harness and settings — lm-evaluation-harness on vLLM, greedy, few-shot completion (no chat template). Percentages are recovery vs. the BF16 original.

Benchmark	BF16 (original)	FP8-dynamic (this repo)	AWQ-8bit	NVFP4A16	NVFP4	AWQ-4bit
MMLU (5-shot)	34.89	35.55 (101.9%)	35.05 (100.5%)	34.84 (99.9%)	35.28 (101.1%)	32.17 (92.2%)
GSM8K (5-shot, strict)	83.70	83.24 (99.5%)	82.71 (98.8%)	82.34 (98.4%)	82.03 (98.0%)	84.00 (100.4%)
ARC-Challenge (0-shot)	53.16	53.24 (100.2%)	53.75 (101.1%)	52.13 (98.1%)	52.99 (99.7%)	52.82 (99.4%)
Winogrande (5-shot)	67.48	66.85 (99.1%)	67.17 (99.5%)	66.30 (98.2%)	66.85 (99.1%)	67.72 (100.4%)
KMMLU (5-shot, direct)	28.46	27.93 (98.1%)	28.40 (99.8%)	29.71 (104.4%)	30.17 (106.0%)	29.73 (104.5%)
HAE-RAE (0-shot)	38.22	37.86 (99.0%)	37.67 (98.6%)	37.31 (97.6%)	38.31 (100.2%)	38.96 (101.9%)
Average	50.98	50.78 (99.6%)	50.79 (99.6%)	50.44 (98.9%)	50.94 (99.9%)	50.90 (99.8%)

Evaluation (measured)

Fidelity to the BF16 original on an internal mixed EN/KO set (100 prompts).

Metric	This model	BF16 original
Perplexity (lower = closer)	1.2376	1.2232 (+1.2%)
Top-1 next-token agreement	96.4%	99.1%

Requirements


Runtime	vLLM — the architecture is in vLLM; tested here with vLLM 0.32.0.
vLLM install	vLLM 0.32 itself needs Python 3.11–3.14, torch 2.13 (CUDA 13.0 build, so an NVIDIA driver with CUDA 13 support), transformers ≥5.16.1,<5.20 and compressed-tensors 0.17.0 (both pinned by vLLM; 0.17.0 reads this checkpoint, which was saved with 0.19.0 — do not upgrade it separately).
GPU	NVIDIA Turing (sm75, e.g. T4) or newer. FP8 math runs on Ada / Hopper / Blackwell (sm89+); on Turing and Ampere vLLM switches to weight-only FP8 automatically. V100 (sm70) is not supported. (Kernel choice per GPU from vLLM 0.32's selection code; only B200 was run here.)
Memory	Weights: 35.5 GB. Add KV cache (grows with context length × concurrent requests) and a few GB for activations / CUDA graphs; use `--tensor-parallel-size N` when the weights do not fit one GPU.
Transformers (without vLLM)	`transformers>=4.56.0` + `compressed-tensors>=0.15.0` (same Llama architecture as MiniCPM5; not run here for Nanbeige).
trust_remote_code	not needed

Usage (vLLM)

Start an OpenAI-compatible server:

vllm serve axteam/Nanbeige4.3-33B-FP8-dynamic --max-model-len 32768

Query it:

curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "axteam/Nanbeige4.3-33B-FP8-dynamic",
  "messages": [{"role": "user", "content": "대한민국의 수도는 어디인가요?"}],
  "max_tokens": 512
}'

Offline inference in Python:

from vllm import LLM, SamplingParams

llm = LLM(model="axteam/Nanbeige4.3-33B-FP8-dynamic", max_model_len=32768)
params = SamplingParams(max_tokens=1024)
out = llm.chat([{"role": "user", "content": "Explain model quantization in two sentences."}], params)
print(out[0].outputs[0].text)

Thinking mode is on by default in the chat template; to turn it off pass chat_template_kwargs={"enable_thinking": False} (Python llm.chat(..., chat_template_kwargs=...), or "chat_template_kwargs": {"enable_thinking": false} in the request body).

Reducing KV-cache memory (optional, serve-time)

The KV cache is kept in BF16 by default. KV-cache precision is a runtime choice, independent of the weight quantization above; for long contexts or large batches FP8 roughly halves KV-cache memory:

vllm serve axteam/Nanbeige4.3-33B-FP8-dynamic --kv-cache-dtype fp8

License

Quantized derivative of Nanbeige/Nanbeige4.3-33B (apache-2.0).

한국어 요약

Nanbeige/Nanbeige4.3-33B의 FP8 (W8A8) 양자화 빌드 · 제작 axteam

개요

항목	내용
원본 모델	`Nanbeige/Nanbeige4.3-33B`
양자화 방식	FP8 동적 양자화 (W8A8)
저장 형식	compressed-tensors `float-quantized` (scheme `FP8_DYNAMIC`)
모델 구조	Dense (Llama 구조) · 52층 · 텍스트 전용
용량	약 35.5 GB
보정 데이터	없음 (가중치 스케일은 가중치만으로 계산, 활성화 스케일은 실행 시 계산)

양자화 상세

가중치: FP8(E4M3), 출력 채널별 스케일
활성화: 실행 시 토큰별 동적 FP8
양자화 절차: PTQ(학습 후 양자화, 재학습 없음), 가중치 스케일은 미리 계산하고 활성화 스케일은 실행 시 토큰마다 계산
BF16 유지: 토큰 임베딩, LM 헤드, 모든 RMSNorm
양자화 대상: 어텐션·MLP 선형층(projection)
원본 파일: 토크나이저·chat template은 원본과 동일

품질 (실측)

원본 대비 충실도: PPL 1.2376 (원본 1.2232, +1.2%, 낮을수록 원본에 가까움), 다음 토큰 top-1 일치율 96.4% (원본 자체 기준값 99.1%) — 내부 영어·한국어 혼합 프롬프트 100개, 원본(BF16)과 비교 측정
벤치마크 평균: 50.78 (원본 50.98, 99.6%) — MMLU·GSM8K·ARC-Challenge·Winogrande·KMMLU·HAE-RAE 6종 평균, lm-evaluation-harness + vLLM, chat template 없는 few-shot·greedy, 모든 빌드를 같은 설정으로 측정
한국어 벤치마크: KMMLU 27.93 (원본 28.46, 98.1%) · HAE-RAE 37.86 (원본 38.22, 99.0%)

요구사항

실행 엔진: vLLM, 이 빌드는 vLLM 0.32.0에서 실행 확인
vLLM 0.32 설치 조건: Python 3.11~3.14, torch 2.13(CUDA 13.0 빌드, CUDA 13 지원 드라이버 필요), transformers 5.16.1 이상 5.20 미만, compressed-tensors 0.17.0 — transformers·compressed-tensors는 vLLM이 버전을 고정하므로 compressed-tensors를 따로 업그레이드하지 말 것 (이 체크포인트는 0.19.0으로 저장됐지만 0.17.0에서도 로드됨)
GPU: NVIDIA Turing(sm75, 예: T4) 이상, V100(sm70) 미지원 — FP8 연산은 Ada·Hopper·Blackwell(sm89 이상)에서 수행, Turing·Ampere는 vLLM이 가중치 전용 FP8로 자동 전환 (GPU별 지원 여부는 vLLM 0.32 커널 선택 코드 기준, 실제 실행은 B200에서만 확인)
메모리: 가중치 35.5 GB + KV 캐시(컨텍스트 길이 × 동시 요청 수에 비례) + 활성화·CUDA graph용 수 GB — GPU 한 장에 다 들어가지 않으면 --tensor-parallel-size N으로 분산
Transformers로 직접 사용: transformers 4.56.0 이상 + compressed-tensors 0.15.0 이상 (MiniCPM5와 같은 Llama 구조, Nanbeige로는 실행 미확인)
trust_remote_code: 불필요

사용법

서버 실행: vllm serve axteam/Nanbeige4.3-33B-FP8-dynamic --max-model-len 32768 — OpenAI 호환 API
Python: LLM(model="axteam/Nanbeige4.3-33B-FP8-dynamic", max_model_len=32768).chat(...) — 전체 예시는 위 영어 본문 'Usage (vLLM)' 절 참고
thinking 모드: 기본 켜짐 — 끄려면 chat_template_kwargs={"enable_thinking": False} 전달(Python은 llm.chat() 인자, HTTP는 요청 본문의 "chat_template_kwargs": {"enable_thinking": false})
KV 캐시 절감(선택): vllm serve에 --kv-cache-dtype fp8 추가 — 기본 BF16인 KV 캐시를 FP8로 저장해 KV 캐시 메모리를 약 절반으로 줄임 (긴 컨텍스트·동시 요청이 많을 때 유리, 가중치 양자화와 별개인 실행 옵션)

라이선스

라이선스: Apache-2.0 — 원본 모델의 라이선스를 그대로 따름
권리: 원본 모델의 권리는 Nanbeige에게 있음, 이 리포는 양자화 파생물

Downloads last month: 25

Safetensors

Model size

33B params

Tensor type

BF16

F8_E4M3

Model tree for axteam/Nanbeige4.3-33B-FP8-dynamic

Base model

Nanbeige/Nanbeige4.3-33B-Base

Finetuned

Nanbeige/Nanbeige4.3-33B

Quantized

(99)

this model