gguf2jns

gguf2jns model.gguf model.jns

gguf2jns converts a GGUF checkpoint into Janas-LLM's JNS file (version 2). Routed experts are rewritten one slot per (layer, expert), with gate, up, and down back to back, so the engine can read an expert with a single aligned direct I/O request. Every other tensor is copied unchanged into the resident region. A dense model has no experts to route: its feed-forward becomes the single slot of each layer, written the same way.A model split into several GGUF files (`name-00001-of-0000N.gguf`) is read from all of them; pass the first file. Data is copied in streaming. The model is never held in memory. The file written is the one the Python converter this tool replaced wrote, byte for byte, so the fingerprints in MODELS.md still hold.janas-get --from-gguf runs this command (then jns_planes when the catalog fingerprint is the plane-cut file). A feed-forward whose width is not a multiple of 256 is widened with zeros to the next multiple, except qwen4exp experts (Qwen3.8-Flash-Next), which keep their width. Gate and up fused in one expert tensor are split into the two matrices of the slot. A resident tensor in MXFP4 is written as Q8_0; routed experts in MXFP4 stay MXFP4.Wrong usage prints a short help and exits 2. Failures print `gguf2jns:` and exit 1. Success prints architecture, layer and expert counts, sizes, copy progress, and `written` with the output size.

model.gguf

Input GGUF. For a split model, the first part (`-00001-of-`). At most 256 parts.

Output JNS path. Created or truncated.

Conversion needs room for both the GGUF and the .jns at once, about twice the model. Keep .jns files on an NVMe SSD: experts are read from disk while the model answers. Only models with routed experts in every layer, or dense ones, are supported. The output is JNS version 2; jns_planes rewrites Q6-K expert downs into version 3 bit planes. Built into `bin/` by Janas's `./build.sh` from `tools/gguf2jns.c`. It is not a separate package.

gguf2jns ships with Janas, written by Maurizio "camauri" Cammalleri (first public release September 2026, GNU GPL 3 or later). It replaced an earlier Python converter and is required to match MODELS.md fingerprints.

janas-get(1), hf2jns_mtp(1), jns_planes(1), janas-chat(1), janas-bench(1), huggingface-cli(1)

Source code · Documentation