bobbydeveloper2014/mamba-siso-1.5b-nimq

Mamba-3 SISO 1.5B

Quantized Mamba-3 SISO checkpoint in Nimformers NIMQ format, paired with its LoRA/Linear++ fine-tune. The base weights use FP4-KM; the LoRA checkpoint is NIMBIN4. The repo slug follows the requested name. Upstream model: .

Contents

• mamba3_siso_1.5b.nimq — FP4-KM base
• lora-g1.bin — 23 R + G LoRA/Linear++ checkpoint
• config.json, tokenizer.json, tokenizer_config.json, special_tokens_map.json — alongside the base model

Architecture: 24 layers, hidden 2048, 64 heads, state 128. The LoRA uses Mamba-3 blocks with a gated Linear++ branch (G) at the final layer. R marks pretrained Mamba-3 blocks.

Download for this repository's inference CLI

From the Nimformers Framework root, download into the expected local layout:

hf download bobbydeveloper2014/mamba-siso-1.5b-nimq \
  --local-dir test/mamba3/models

Then run make -f test/mamba3/Makefile infer and use the checkpoint path test/mamba3/models/lora-g1.bin. The tokenizer sidecar and model config are included for that local layout.

This repository contains the locally supplied 1.5B SISO base checkpoint.

Tokenizer and config files are stored at the repo root beside the model file. The local tokenizer sidecar points to test/mamba3/models.

Framework

Use this model with Nimformers Framework.