You can deploy this model and start earning money today!

bong-water-water-bong/
BlackMamba-1.5B-1BP

Name: BlackMamba-1.5B-1BP
Availability: InStock

1bp

q4nx

mamba

Zyphra/BlackMamba-1.5B

finetune

apache-2.0

Views

No views yet

BlackMamba-1.5B — 1BP format

Zyphra/BlackMamba-1.5B (Apache-2.0) converted to 1BP — the single-file model format used by 1bit.systems's inference engine.

1BP packs everything a loader needs into one memory-mappable file: a 256-byte header with model config, a variable-length tensor index, and Q4NX-tiled (32x256, 4-bit groups of 32 with bf16 scale/zero-point) weight data.

Architecture: Mamba1 (not Mamba2) + top-1 Sinkhorn-routed MoE, alternating per layer per the base checkpoint's mamba_moe_layers pattern — every other layer is a pure Mamba1 SSM block, the rest are pure MoE SwiGLU FFN blocks (never both in the same layer). 30 layers, hidden size 1152, 8 experts per MoE layer.

Conversion notes — three real bugs found and fixed

scripts/blackmamba_to_gguf.py (standalone BlackMamba → GGUF converter, since no upstream GGUF export exists for this architecture) shipped with three genuine correctness bugs, caught by cross-checking a new CPU reference tool (tools/blackmamba_cpu_reference.cpp) against a real PyTorch forward pass using the official reference implementation (github.com/Zyphra/BlackMamba):

Q4_0 nibble encoding was wrong. The 4-bit values were packed as two's-complement instead of the excess-8/offset-binary encoding the actual dequantizer expects (nibble - 8 = value, not value & 0xF) — this silently scrambled nearly every quantized weight matrix in the model. The shared reader's own code has a comment noting this exact bug pattern already bit this codebase once before in different code.
conv1d weight .reshape() misuse — reshape reinterprets a flat buffer with new dimensions, it does not transpose; the wrong target shape silently scrambled which (channel, kernel-tap) pair each value belonged to.
MoE router bias silently dropped, and the router weight was being quantized to 4-bit despite being an 8-way argmax-routing matrix, not a dense weight (now kept F32 in the source GGUF).

All three are fixed as of this upload. Verified: layer-0 SSM output now matches a real PyTorch forward pass to within 0.04% (the residual difference is expected 4-bit quantization noise, not a bug) — structural tensor count matches the source GGUF exactly (438 tensors), and full-sequence decode produces healthy, non-repeating token distributions matching the pattern from a real end-to-end run of the official reference implementation.

Real measured numbers

GPU — official Zyphra reference implementation, full precision, PyTorch/ROCm on Radeon 8060S. Median of 5 runs per prompt length, 128-token greedy decode each:

Prompt	Prompt len	Prefill	Decode	Spread
Short	2	24.4 ms/tok	42.7 tok/s	42.6–46.4
Medium	10	23.7 ms/tok	42.8 tok/s	42.6–42.8
Long	30	23.5 ms/tok	42.4 tok/s	42.2–42.5

CPU — tools/blackmamba_cpu_reference.cpp, F32, scalar, single-threaded, reading this repo's GGUF directly. Median of 3 runs, 60-token decode:

8.5 tok/s (117.4–119.4 ms/tok across runs)

Neither is the Q4NX-quantized 1BP file's actual production speed — no fast inference path exists for this architecture in the engine yet (see Usage below). These are correctness/reference baselines, not production numbers.

Usage

No fast inference path exists for this architecture in the engine yet. engine/npu/'s npu_engine_universal/onebp_loader.cpp has no Mamba/SSM support, and the Mamba1 GPU kernels that do exist (src/mamba1_engine.hip) aren't wired into any build target or loader that understands this checkpoint's MoE-expert tensor layout. Until that's built, the only way to actually run this model in this repo is the reference implementation used to produce the numbers above:

bash

git clone https://github.com/bong-water-water-bong/1bit-systems
cd 1bit-systems
g++ -std=c++23 -O2 -Iinclude -Isrc tools/blackmamba_cpu_reference.cpp src/gguf_reader.cpp -o blackmamba_cpu_reference
# Note: reads the GGUF directly (F32 dequant), not this repo's .1bp file --
# the 1BP loader doesn't have Mamba support wired in yet either.
python3 scripts/blackmamba_to_gguf.py Zyphra/BlackMamba-1.5B blackmamba-1.5b.gguf
./blackmamba_cpu_reference blackmamba-1.5b.gguf 20

Attribution

Base model: Zyphra/BlackMamba-1.5B, Apache-2.0. This repo contains a re-quantized (Q4NX, 4-bit) derivative in a different on-disk format; the original license terms apply to the underlying weights.

bong-water-water-bong/BlackMamba-1.5B-1BP