Official implementation of BitMamba-2. A scalable 1.58-bit State Space Model (Mamba-2 + BitNet) trained from scratch on 150B tokens. Includes JAX training code and high-performance C++ inference engine.

Portable C inference engine for BitMamba-2 models

Ultra-lightweight C++ inference engine for BitMamba-2 (1.58-bit SSM). Runs 1B models on consumer CPUs at 50+ tok/s using <700MB RAM. No heavy dependencies.

Agentic-1.58b: A BitMamba reasoning engine built for consumer GPUs. By fusing 1.58-bit ternary quantization with Mamba-2 State Space Models via custom Triton kernels, this pipeline achieves massive context scaling on a single RTX 3090. Includes full scripts for pre-training, SFT, and GRPO reinforcement learning.