Ultra memory-efficient LLM inference engine.

ZSE is designed to run large language models with minimal memory footprint while maintaining high performance. Our key innovation is the Intelligence Orchestrator that provides smart recommendations based on your available (not total) memory.

• 🧠 zAttention: Custom CUDA kernels for paged, flash, and sparse attention
• 🗜️ zQuantize: Per-tensor INT2-8 mixed precision quantization
• 💾 zKV: Quantized KV cache with sliding precision (4x memory savings)
• 🌊 zStream: Layer streaming with async prefetch (run 70B on 24GB GPU)
• 🎯 zOrchestrator: Smart recommendations based on FREE memory
• 📊 Efficiency Modes: speed / balanced / memory / ultra

3.9s (7B) and 21.4s (32B) to first token with .zse format — verified on A100-80GB.

# One-time conversion (~20s)
zse convert Qwen/Qwen2.5-Coder-7B-Instruct -o qwen-7b.zse

# Every subsequent start: 3.9s
zse serve qwen-7b.zse

Note: Results measured on A100-80GB with NVMe storage (Feb 2026). On consumer SSDs expect 5-10s; HDDs may be slower. Any modern SSD achieves sub-10s cold starts.

32B note: Use NF4 (19.3 GB) on GPUs with <36 GB VRAM. Use .zse (35 GB, 5.6× faster start) on 40 GB+ GPUs.

With CUDA support (recommended):

pip install zllm-zse[cuda]

From source:

git clone https://github.com/Zyora-Dev/zse.git
cd zse
pip install -e ".[dev]"

# Any HuggingFace model works!
zse serve Qwen/Qwen2.5-7B-Instruct
zse serve meta-llama/Llama-3.1-8B-Instruct
zse serve mistralai/Mistral-7B-Instruct-v0.3
zse serve microsoft/Phi-3-mini-4k-instruct
zse serve google/gemma-2-9b-it

# With memory optimization
zse serve Qwen/Qwen2.5-32B-Instruct --max-memory 24GB

# With recommendations
zse serve meta-llama/Llama-3.1-70B-Instruct --recommend

# Ultra memory efficiency
zse serve deepseek-ai/DeepSeek-V2-Lite --efficiency ultra

# GGUF models (via llama.cpp)
zse serve ./model-Q4_K_M.gguf

💡 Supported Models: Any HuggingFace transformers model, safetensors, GGUF, or .zse format. Popular choices: Qwen, Llama, Mistral, Phi, Gemma, DeepSeek, Yi, and more.

zse chat Qwen/Qwen2.5-7B-Instruct

zse convert Qwen/Qwen2.5-32B-Instruct -o qwen-32b.zse --target-memory 24GB

ZSE provides an OpenAI-compatible API:

zse serve Qwen/Qwen2.5-7B-Instruct --port 8000

import openai

client = openai.OpenAI(base_url="http://localhost:8000/v1", api_key="zse")

response = client.chat.completions.create(
    model="Qwen/Qwen2.5-7B-Instruct",
    messages=[{"role": "user", "content": "Hello!"}]
)
print(response.choices[0].message.content)

zse serve model --efficiency memory

zse serve model --mode dev

• No authentication required
• SQLite database
• Hot reload enabled
• Debug logging

zse serve model --config configs/enterprise.yaml

• API key authentication
• PostgreSQL + Redis
• Prometheus metrics
• Rate limiting
• Multi-tenancy

zse/
├── core/                   # ZSE Native Engine (100% custom)
│   ├── zattention/         # Custom attention kernels
│   ├── zquantize/          # Quantization (GPTQ, HQQ, INT2-8)
│   ├── zkv/                # Paged + quantized KV cache
│   ├── zstream/            # Layer streaming + prefetch
│   ├── zscheduler/         # Continuous batching
│   └── zdistributed/       # Tensor/pipeline parallelism
├── models/                 # Model loaders + architectures
├── engine/                 # Executor + Orchestrator
├── api/                    # CLI, FastAPI server, Web UI
└── enterprise/             # Auth, monitoring, scaling

GGUF models are supported via llama.cpp backend:

pip install zllm-zse[gguf]
zse serve ./model.gguf

Note: GGUF uses llama.cpp for inference. Native ZSE engine handles HuggingFace, safetensors, and .zse formats.

# CPU
docker run -p 8000:8000 ghcr.io/zyora-dev/zse:latest

# GPU (NVIDIA)
docker run --gpus all -p 8000:8000 ghcr.io/zyora-dev/zse:gpu

# With model pre-loaded
docker run -p 8000:8000 -e ZSE_MODEL=Qwen/Qwen2.5-0.5B-Instruct ghcr.io/zyora-dev/zse:latest

Docker Compose:

docker-compose up -d                    # CPU
docker-compose --profile gpu up -d      # GPU

See deploy/DEPLOY.md for full deployment guide including Runpod, Vast.ai, Railway, Render, and Kubernetes.

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Run tests with coverage
pytest --cov=zse

# Type checking
mypy zse

# Linting
ruff check zse

Apache 2.0

• PagedAttention concept from vLLM (UC Berkeley)
• Flash Attention from Tri Dao
• GPTQ, HQQ, and other quantization research