Convert supported checkpoints into NInfer artifacts using a guided wizard or manual controls. A standalone Windows x64 desktop application and JSON backend share the same conversion engine.
Using NInferEZ Manager? Its Converter page provides the same workflow without opening another application.
Download 1.0.0 · Supported inputs · Backend API · Engine · Models
Extract the complete Portable ZIP into a writable folder and open NInferEZ.Converter.exe. End users do not need a separate Python, .NET or CUDA Toolkit installation.
The bundled conversion runtime is CPU-only and works offline on local sources. Conversion still consumes CPU, RAM and disk. Optional inference-engine testing uses that engine's GPU and is explicitly selected. The standalone release is Portable; no separate Converter installer is included.
Guided wizard: select a source, attach matching resources, inspect available options, choose a prepared plan and review the destination. Missing files and unsupported choices are explained before conversion.
Manual setup: profiles, components, source mappings, recipes, tensor policies, chunk/shard limits and optional engine validation. Request import/export preserves advanced options.
Both modes use the existing backend with progress and cancellation. Changed sources or plans require a fresh compatibility check. Custom Python recipes execute code with the user's permissions; only use recipes you trust.
Support follows source geometry and importer coverage, not a filename alone:
- Compatible Qwen3.5 Dense/MoE geometry, including supported Qwen3.6/3.8 derivatives.
- Safetensors checkpoints/indexed shards with matching configuration, tokenizer and frontend resources.
- Supported compressed sources, including native NVFP4 and row-/tensor-scaled FP8 layouts.
- Supported GGUF encodings, accompanied by matching HF configuration/tokenizer resources.
- Native v3 repacking and qualified older-container import paths.
- Optional vision, MTP and DFlash/DFlash2 when compatible companion sources are supplied.
See supported inputs and the backend's prepared plans for exact choices. Output-format support does not imply an importer exists for every architecture. Unsupported mappings are rejected rather than guessed.
Higher-bit output cannot restore information already lost in compressed sources. Output-size estimates are not quality scores, speed predictions or total VRAM estimates.
.\NInferEZ.Converter.Backend.exe capabilities
.\NInferEZ.Converter.Backend.exe probe --request check.json
.\NInferEZ.Converter.Backend.exe analyze --request check.json
.\NInferEZ.Converter.Backend.exe convert --request convert.json
.\NInferEZ.Converter.Backend.exe validate --request validate.jsonSee BACKEND.md for schemas and NDJSON events. The EXE locates its runtime relative to itself; callers need not use a particular working directory.
Structural validation uses the artifact reader and a host-only engine inspector. It is distinct from GPU execution and answer-quality evaluation. See QA.
Building requires .NET SDK 10, and Python 3.12 plus uv when preparing a new runtime. Pass -PythonSource explicitly to scripts/build.ps1 -PrepareRuntime; the build script contains no personal default path. scripts/package.py writes manifests, checksums and the archive.
Application additions and the engine-derived converter use Apache-2.0. The vendored converter is pinned to Engine revision 51cdf17108d28d7b772dfcc56a80dcf5de517291; vendor changes record adaptations and retain upstream authorship.
Packages include Python, PyTorch, NumPy, GGUF, .NET, native-inspector and runtime licenses/notices. Converted weights retain their original model licenses. NInferEZ is an independent distribution.