Instructions to use axteam/Motif-2.6B-FP8-dynamic with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use axteam/Motif-2.6B-FP8-dynamic with Transformers:

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

pipe = pipeline("text-generation", model="axteam/Motif-2.6B-FP8-dynamic", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("axteam/Motif-2.6B-FP8-dynamic", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use axteam/Motif-2.6B-FP8-dynamic with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "axteam/Motif-2.6B-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/Motif-2.6B-FP8-dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/axteam/Motif-2.6B-FP8-dynamic

SGLang

How to use axteam/Motif-2.6B-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/Motif-2.6B-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/Motif-2.6B-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/Motif-2.6B-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/Motif-2.6B-FP8-dynamic",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use axteam/Motif-2.6B-FP8-dynamic with Docker Model Runner:
```
docker model run hf.co/axteam/Motif-2.6B-FP8-dynamic
```

Motif-2.6B · FP8 (dynamic)

8-bit quantization of Motif-Technologies/Motif-2.6B for 🤗 Transformers (trust_remote_code). Quantized by axteam.


Base model	`Motif-Technologies/Motif-2.6B`
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 (Motif arch · differential attention) · 32 layers · text-only
Size on disk	~3.9 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.

Requirements


Runtime	Transformers only — vLLM 0.32 does not support the Motif architecture.
Transformers	`transformers>=4.56,<5` + `compressed-tensors>=0.19` + `trust_remote_code=True` (tested on CPU: 4.57.6 + 0.20.1, 4.56.2 + 0.20.1, 4.57.6 + 0.19.0). Transformers 5.x does not work.
Memory	weights 3.9 GB, but Transformers decompresses them to BF16 (~5.2 GB) on the first forward.

Usage (Transformers)

vLLM does not support the Motif architecture (checked with vLLM 0.32). The bundled modeling_motif.py is the upstream code written for Transformers 4.x — under Transformers 5.x it fails to load or generates only ! tokens (the BF16 original behaves the same), so pin 4.x:

pip install "transformers>=4.56,<5" "compressed-tensors>=0.19" accelerate torch

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "axteam/Motif-2.6B-FP8-dynamic"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16).eval()

msgs = [{"role": "user", "content": "대한민국의 수도는 어디인가요?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, attention_mask=torch.ones_like(ids), max_new_tokens=128, do_sample=False,
                     pad_token_id=tok.pad_token_id)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Tested on CPU, 2026-11-01: transformers 4.57.6 + compressed-tensors 0.20.1 (full prompt set), 4.56.2 + 0.20.1 and 4.57.6 + 0.19.0 (quick check); torch 2.15, Python 3.12. With the chat prompt above it answers 대한민국의 수도는 서울입니다. …, token-for-token the same as the BF16 original; the first next-token probabilities after "The capital of France is" are Paris 0.59 (FP8) vs 0.62 (BF16). Transformers 4.51.3 + compressed-tensors 0.10.2 also works but is very slow (it re-applies the FP8 quantization on every forward). 4.53.3 loads but crashes on the first forward (the FP8 weights are never decompressed — Transformers' compressed-tensors loader decompresses them only from 4.56); 4.51.3 with compressed-tensors ≥0.11 does not load. Memory/speed: Transformers decompresses the FP8 weights to BF16 on the first forward, so runtime memory and speed are the same as the BF16 model; this build mainly reduces download and storage size.

License

Quantized derivative of Motif-Technologies/Motif-2.6B (apache-2.0).

한국어 요약

Motif-Technologies/Motif-2.6B의 FP8 (W8A8) 양자화 빌드 · 제작 axteam

개요

항목	내용
원본 모델	`Motif-Technologies/Motif-2.6B`
양자화 방식	FP8 동적 양자화 (W8A8)
저장 형식	compressed-tensors `float-quantized` (scheme `FP8_DYNAMIC`)
모델 구조	Dense (Motif 구조, differential attention) · 32층 · 텍스트 전용
용량	약 3.9 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)

원본과 비교 (CPU, 2026-11-01): 채팅 프롬프트 답이 BF16 원본과 토큰 단위로 동일 (대한민국의 수도는 서울입니다. …), "The capital of France is" 다음 토큰 Paris 확률 0.59 (원본 0.62) — transformers 4.57.6 + compressed-tensors 0.20.1

요구사항

실행 환경: Transformers 전용 — 공식 vLLM 0.32는 Motif 구조 미지원
Transformers: transformers 4.56 이상 5 미만 + compressed-tensors 0.19 이상 + trust_remote_code=True (CPU에서 확인: 4.57.6 + 0.20.1, 4.56.2 + 0.20.1, 4.57.6 + 0.19.0), transformers 5.x는 동작하지 않음
메모리: 가중치 3.9 GB — 첫 실행 때 BF16(약 5.2 GB)으로 풀려 메모리·속도는 원본과 같으므로 주된 효과는 다운로드·저장 용량 절감

사용법

설치: pip install "transformers>=4.56,<5" "compressed-tensors>=0.19" accelerate torch
로드: AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16) → chat template 적용 후 generate (전체 예시는 위 영어 본문 'Usage' 절 참고)

라이선스

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

Downloads last month: 151

Safetensors

Model size

3B params

Tensor type

F32

F8_E4M3

Model tree for axteam/Motif-2.6B-FP8-dynamic

Base model

Motif-Technologies/Motif-2.6B

Quantized

(2)

this model