Downloads · 30 days
0
bong-water-water-bong/BlackMamba-2.8B-1BP
BlackMamba-2.8B-1BP is a machine learning model from bong-water-water-bong. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Zyphra/BlackMamba-2.8B (Apache-2.0) converted to 1BP — the single-file model format used by 1bit.systems's inference engine.
Downloads · 30 days
0
Access
Public
Updated Jul 23, 2026
Repo size
5.6 GB
Likes
0
Public
Click a slice to open those files.
.1bp1.9 GB · 100%
From the Hugging Face model README
Zyphra/BlackMamba-2.8B (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). 36 layers, hidden size 1472, 8 experts per MoE layer.
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):
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..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.All three are fixed as of this upload. Verified via the same fix set applied and cross-checked on the 1.5B sibling model (layer-0 SSM output matched a real PyTorch forward pass to within 0.04%, expected 4-bit quantization noise). Structural tensor count matches the source GGUF exactly (525 tensors), and full-sequence decode produces healthy, mostly-non-repeating token distributions.
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 | 36.9 ms/tok | 27.8 tok/s | 27.6–28.6 |
| Medium | 10 | 36.1 ms/tok | 27.8 tok/s | 27.7–27.9 |
| Long | 30 | 36.2 ms/tok | 28.1 tok/s | 27.7–28.6 |
CPU — tools/blackmamba_cpu_reference.cpp, F32, scalar, single-threaded, reading this repo's GGUF directly. Median of 3 runs, 60-token decode:
4.6 tok/s (210.9–216.2 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.
Note: at 128-token greedy decode on the arbitrary fixed test prompt, the "medium" and "long" runs fall into short repetition loops after the initial tokens (4/128 and 8/128 unique tokens respectively) — a well-known general characteristic of pure greedy decoding without repetition penalty on any model, not specific to this conversion. The short 15-token runs used for the initial correctness check (and the 1.5B sibling model at this same length) stayed fully non-degenerate.
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.
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:
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-2.8B blackmamba-2.8b.gguf
./blackmamba_cpu_reference blackmamba-2.8b.gguf 20
Base model: Zyphra/BlackMamba-2.8B, 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.