README.md

metadata

title: 88plug AI Lab
emoji: 🔌
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false

88plug AI Lab

Production-grade compressed-tensors quantizations of frontier LLMs, VLMs, and omni models, engineered for native vLLM v0.9.0+ deployment. Every model is validated against the baseline on MMLU and ships with a complete vLLM-ready configuration.

---

Why compressed-tensors

Most quantization formats (AWQ, GPTQ, GGUF) target a single inference backend and ship a frozen weight layout that cannot be further composed or modified at load time. compressed-tensors is the format developed by Neural Magic and maintained as a first-class vLLM citizen. Key differences:

• Native vLLM integration. No format conversion, no plugin shims. vLLM reads compressed-tensors models directly via its built-in CompressedTensorsWorker. This means full PagedAttention, continuous batching, and tensor parallelism work without modification.
• Composable precision. A single checkpoint can carry per-layer or per-group precision assignments. Mixed-precision MoE configurations (e.g., FP8 attention + INT4 experts) are expressed in the same file, not hacked around.
• Reproducible calibration metadata. The quantization config, calibration scheme, and per-channel scales are stored inside the checkpoint. What you see in the config is exactly what ran.
• Forward compatibility. As vLLM adds new kernel support (FP8, INT8, sparse), compressed-tensors models gain that support without re-quantizing.

AWQ and GPTQ remain fine for llama.cpp and older toolchains. If you are deploying on vLLM in production, compressed-tensors is the correct choice.

---

Quality Standard

All models are quantized with AutoRound (iters=200) or RTN where noted.

Tier	Method	Target Recovery	Hardware Floor
W8A16	RTN / AutoRound iters=200	Near-lossless (>99.5% MMLU)	Ampere (A100, A6000, RTX 30xx+)
W4A16	AutoRound iters=200	≥99% MMLU vs FP16 baseline	Ampere (A100, A6000, RTX 30xx+)

AutoRound at iters=200 runs sign-gradient optimization over a calibration set to minimize weight rounding error. At W4A16, this closes most of the gap between naive round-to-nearest and GPTQ/AWQ, while producing a checkpoint that vLLM can load natively.

---

Model Catalog

All 16 models are in compressed-tensors format, validated for vLLM v0.9.0+.

Qwen3.6-35B-A3B — Mixed-Precision MoE, 1M context

Precision	Repo	Architecture
W8A16	88plug/Qwen3.6-35B-A3B-W8A16	MoE, 35B total / 3.6B active
W4A16	88plug/Qwen3.6-35B-A3B-W4A16	MoE, 35B total / 3.6B active

Qwen3.6-27B — Dense Hybrid, 262k context

Precision	Repo	Architecture
W8A16	88plug/Qwen3.6-27B-W8A16	Dense, 27B
W4A16	88plug/Qwen3.6-27B-W4A16	Dense, 27B

Qwen3-Omni-30B-A3B — Audio + Vision + Speech

Precision	Repo	Architecture
W8A16	88plug/Qwen3-Omni-30B-A3B-W8A16	Omni MoE, 30B / 3B active
W4A16	88plug/Qwen3-Omni-30B-A3B-W4A16	Omni MoE, 30B / 3B active

Qwen2.5-Omni-7B — Efficient Omni

Precision	Repo	Architecture
W8A16	88plug/Qwen2.5-Omni-7B-W8A16	Omni dense, 7B
W4A16	88plug/Qwen2.5-Omni-7B-W4A16	Omni dense, 7B

Gemma4-E4B-it — Vision-Language Model

Precision	Repo	Architecture
W8A16	88plug/Gemma4-E4B-it-W8A16	VLM, 4B
W4A16	88plug/Gemma4-E4B-it-W4A16	VLM, 4B

Gemma4-E2B-it — Ultra-Efficient VLM

Precision	Repo	Architecture
W8A16	88plug/Gemma4-E2B-it-W8A16	VLM, 2B
W4A16	88plug/Gemma4-E2B-it-W4A16	VLM, 2B

MiniCPM-o-4.5 — Omni Model

Precision	Repo	Architecture
W8A16	88plug/MiniCPM-o-4.5-W8A16	Omni dense
W4A16	88plug/MiniCPM-o-4.5-W4A16	Omni dense

Nemotron-3-Nano-30B-A3B — Hybrid SSM/Attention

Precision	Repo	Architecture
W8A16	88plug/Nemotron-3-Nano-30B-A3B-W8A16	Hybrid SSM/Attention MoE
W4A16	88plug/Nemotron-3-Nano-30B-A3B-W4A16	Hybrid SSM/Attention MoE

---

Quickstart

Requires vLLM v0.9.0+ and an Ampere-class GPU (A100, A6000, RTX 3090/4090, or equivalent).

Install

pip install vllm>=0.9.0

Launch (offline inference)

from vllm import LLM, SamplingParams

llm = LLM(
    model="88plug/Qwen3.6-35B-A3B-W4A16",
    max_model_len=131072,          # adjust to available VRAM
    tensor_parallel_size=1,        # increase for multi-GPU
)

sampling_params = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=512)

outputs = llm.generate(
    ["Explain the tradeoffs between W4A16 and W8A16 quantization for production inference."],
    sampling_params,
)

print(outputs[0].outputs[0].text)

Launch (OpenAI-compatible server)

vllm serve 88plug/Qwen3.6-35B-A3B-W4A16 \
    --max-model-len 131072 \
    --tensor-parallel-size 1 \
    --port 8000

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "88plug/Qwen3.6-35B-A3B-W4A16",
        "messages": [{"role": "user", "content": "What is compressed-tensors?"}],
        "max_tokens": 256
    }'

---

Hardware Requirements

Model Size	W8A16 VRAM	W4A16 VRAM	Recommended
2B–7B	8–16 GB	6–10 GB	Single A6000 / RTX 4090
27B–35B (dense)	32–40 GB	20–28 GB	Single A100 80G or 2x A6000
30B–35B (MoE, 3B active)	28–36 GB	18–24 GB	Single A100 80G or 2x A6000

Active-parameter MoE models load all expert weights into VRAM but only route through a subset per token. VRAM requirement is determined by total parameters, not active parameters.

---

Contact

Developer: Andrew Mello
Organization: huggingface.co/88plug
Issues and model requests: open a discussion on the relevant model repo.

Model uploads are automated via the 88plug-bot account.