Intern-S2-Mobius-4bit-mlx-mtp

Intern-S2-Mobius MTP head (4-bit MLX)

Companion repo — not a standalone model. This holds the Multi-Token-Prediction (MTP) head for nightscape/Intern-S2-Mobius-4bit-mlx, the base 4-bit MLX conversion of internlm/Intern-S2-Mobius.

Files:

• model-mtp-head.safetensors — the MTP head weights (~463 MB).
• config.json — the checkpoint config with text_config.mtp_num_hidden_layers: 1 and the mtp-layer quantization entries (mtp gate at 8-bit).

The base model's 5 weight shards, tokenizer, and remote model code are not duplicated here — download the base repo and overlay this head + config alongside them.

Assemble (omlx)

omlx loads an MTP-enabled model as base weights plus this head:

# put the base repo on disk, then drop the head + config next to it:
cp model-mtp-head.safetensors config.json <base-dir>/

Then serve with the omlx build that ships the interns2_mobius MTP driver (nightscape/omlx, add-interns2-mobius).

Notes

• MTP is experimental: ~1.9x lossless decode on the M4 Max, but exclusive to the omlx interns2_mobius MTP driver. Stock mlx-lm does not load this head.
• --trust-remote-code is required (custom InternS1 tokenizer + interns2_mobius code).

License

Apache-2.0 — see LICENSE. Model by InternLM; MTP head weights from the internlm/Intern-S2-Mobius checkpoint, conversion and MLX port by this repo's authors.