---

license: apache-2.0

base_model:

• Zyphra/BlackMamba-2.8B

tags:

• gguf
• mamba
• moe

---

BlackMamba-2.8B — GGUF

Our own GGUF conversion of Zyphra's BlackMamba-2.8B: Mamba-1 layers alternating with a Switch MoE (sigmoid top-1 router with bias, gated GELU experts). The model code is ours (src/models/blackmamba.cpp in our llama.cpp fork); upstream llama.cpp has no BlackMamba.

Contents

• BlackMamba-2.8B-Q8_0.gguf

Validation

• The same port matches Zyphra's own PyTorch code (FP32, CPU) at 96/96 teacher-forced positions on BlackMamba-1.5B. The 2.8B was not separately teacher-forced.
• Wikitext-2 test perplexity, 60 chunks of 512 tokens, Q8_0 on Vulkan: 15.28 ± 0.34 (BlackMamba-1.5B: 17.84 on the same test).
• Decode on Strix Halo (Radeon 8060S, Vulkan), Q8_0: 229 tok/s.
• It is a base model: it continues text rather than chatting.

Running it

With the 1bit engine:

1bit serve -m BlackMamba-2.8B-Q8_0.gguf --device vulkan

BlackMamba runs from our llama.cpp fork (branch 1bit/vulkan-upstream), which also has a Vulkan Mamba-1 scan; upstream llama.cpp has neither.

Attribution

• Base model: Zyphra/BlackMamba-2.8B, Apache 2.0.
• GGUF conversion and validation: the 1bit engine project, with the converter and model code in our llama.cpp fork.
• License: Apache 2.0, inherited from the base model.