Nemotron 3.5 Lightning 30B-A3B

In development

Status: in development. This entry is not in any released or TestFlight build of Local Mind, and no release date is promised. It is documented here because what has been verified about it is worth recording precisely.

A 30-billion-parameter mixture-of-experts model from NVIDIA, published by EigenLabs as a 4-bit MLX pack. Local Mind catalogues that pack.

What the catalogue records

Format	MLX checkpoint directory, ten files
Total size	18,544,223,673 bytes (about 18.5 GB)
Quantisation	4-bit affine, group size 64
Extra head	embedded multi-token prediction, one predicted layer
Trained context	262,144 tokens
Licence	OpenMDW-1.1

Every one of the ten files is catalogued with its size and SHA-256 digest, taken from the publisher's repository at a pinned revision.

What has been verified

The pack was downloaded and checked end to end:

All ten digests matched, and their sizes sum to exactly the recorded total.
A stock MLX server loaded the directory and began serving. Of the pack's 763 tensors, 34 belong to the multi-token-prediction head, and these are tolerated rather than rejected as unexpected keys.
It answered a prompt correctly, with its reasoning content separated from the answer — which also confirms that the reasoning format is derived correctly from the chat template recorded in the catalogue.
The required pre-weight file set was measured rather than assumed: given a directory holding only the configuration file, the server's first complaint is about a missing weight tensor, so no auxiliary transform file is needed. Some packs do require one, and that failure mode would have appeared here.

What is not claimed

No decode or prefill speed. The one run to date was a single request, on a machine concurrently serving two other models, with the weights on an external USB volume. Those conditions move throughput by far more than any difference worth reporting, so the entry keeps its unevaluated badge and every speed shown for it is modelled from the model's structure rather than measured.

A measurement will replace this section when one exists in the shipped configuration, on a quiet machine, with the weights on internal storage.

Engine

The model declares a tuned MLX engine by name; see Inference engines. Performance work for this architecture takes place in a public optimisation track, and no standing or score in that track is claimed here.