---

license: apache-2.0

base_model:

• Zyphra/BlackMamba-1.5B

tags:

• gguf
• mamba
• moe

---

BlackMamba-1.5B — GGUF

Our own GGUF conversion of Zyphra's BlackMamba-1.5B

(a hybrid Mamba/MoE architecture), added to the 1bit engine's

architecture showcase. From-scratch conversion and a new src/models/blackmamba.cpp in

our llama.cpp fork — BlackMamba wasn't previously supported by llama.cpp.

Contents

• BlackMamba-1.5B-Q8_0.gguf

Validation

96/96 teacher-forced token agreement against Zyphra's own PyTorch reference implementation

(run on CPU with Triton/CUDA kernels stubbed out).

A few non-obvious things the config doesn't tell you, that our conversion had to get

right: the config says swiglu but the activation is actually gated erf-GELU

(F.gelu); the router's Linear has a bias; the final norm is a biased LayerNorm,

not RMSNorm; the checkpoint's embedding key is embedding.word_embeddings.weight; and

config.json doesn't declare an architectures field at all.

Measured performance (Strix Halo, Vulkan, Q8_0)

pp512: 12890 tok/s · tg128: 378 tok/s

Running it

1bit serve -m BlackMamba-1.5B-Q8_0.gguf --device vulkan

Attribution

• Base model: Zyphra/BlackMamba-1.5B, Apache 2.0.
• GGUF conversion and llama.cpp architecture support: this project's engine team.
• License: Apache 2.0, inherited from the base model.