Choose a source.
Select a compatible local checkpoint, supported GGUF, or a qualified NInfer container.
Prepare supported models for NInfer with a guided desktop workflow. Start in Manager, or use the independent Windows Converter.
Windows x64 · CPU-only conversion · Portable 1.0.0
PREPARE · INSPECT · CONVERTSelect a compatible local checkpoint, supported GGUF, or a qualified NInfer container.
Match the configuration, tokenizer, and optional components required by your source.
Inspect supported profiles and output choices. Compatibility checks explain missing files.
Follow progress, cancel when needed, and inspect the resulting NInfer artifact.
Guided mode walks through the source, compatibility, plan, and destination. Manual mode exposes profiles, components, mappings, tensor policies, and chunk or shard limits.
Both use the same backend. Import and export requests to keep advanced plans reusable.
Compatibility comes from source geometry, tensor encodings, and importer coverage.
Compatible Qwen3.5 Dense/MoE geometry, including supported Qwen3.6 and Qwen3.8 derivatives. Safetensors checkpoints and indexed shards need matching resources.
Supported native NVFP4, row- or tensor-scaled FP8, and supported GGUF encodings. Higher-bit output cannot restore information lost in compression.
Vision, MTP, and DFlash/DFlash2 when compatible companion sources are provided. Structural checks are separate from GPU execution or quality evaluation.
The standalone Portable includes its CPU conversion runtime and works with offline local sources. Conversion uses CPU, RAM, and disk. Optional GPU engine testing is selected explicitly.
Your models, your workspace, your next idea.
Start with NInferEZ Manager.