SipsaLabs/mamba2-2.7b-uc-v3-bpw5

Mamba-2 2.7B -- UltraCompress v3 (5-bit, near-lossless)

Compressed by: Sipsa Labs Codec: UltraCompress v0.6+ (BUSL-1.1) Base model: state-spaces/mamba2-2.7b

Headline numbers

Metric	Value
PPL ratio (vs bf16 baseline)	1.0133x
Baseline PPL (bf16)	25.028
Compressed PPL (5-bit)	25.362
Effective bpw	~5.1
Architecture class	SSM (Mamba-2)
Peak VRAM (single-GPU eval)	~12.9 GB
Verified matrix	cross-architecture (dense LLM, MoE, SSM, ViT)
Verifier hash	SHA256_MANIFEST.json

PPL is measured on a held-out FineWeb-edu corpus, n=30 prompts of 1024 tokens each, against the bf16 reference. Methodology details and reproducible measurement scripts ship with the verifier -- see below.

This is the second state-space model (SSM) architecture in the Sipsa Labs verified matrix, following Mamba-2.8B. Together they demonstrate cross-architecture generalization from Mamba-1 to Mamba-2 with zero format changes.

Important: upstream config.json fix required

The upstream state-spaces/mamba2-2.7b checkpoint ships with an incorrect config.json. Four fields must be corrected at load time for HF Transformers Mamba2ForCausalLM to produce valid outputs:

Field	Upstream (wrong)	Correct
hidden_size	4096	2560
num_heads	128	80
state_size	128	16
vocab_size	50277	50288

Additionally, the embedding key is mismatched (backbone.embedding.weight in the checkpoint vs backbone.embeddings.weight in the model). Our evaluation scripts apply these corrections automatically. No mamba_ssm library install is required -- the HF Transformers naive SSM fallback path handles inference.

Comparator note

This pack uses an SSM-compatible comparator -- the same architecture-matched reconstruction comparator used for the Mamba-1 record. The transformer comparator pipeline (used for transformer LLM records in the registry) is architecture-incompatible with SSMs. Treat this record as the canonical SSM near-lossless 5-bit number; do not directly compare against transformer records without reading this note.

Verify the reconstruction yourself

pip install ultracompress
uc verify SipsaLabs/mamba2-2.7b-uc-v3-bpw5
# SHA-256 match: True
# PPL ratio (vs base): 1.0133x

The verifier downloads this artifact, decompresses to an HF folder, and checks that every weight tensor is np.allclose to the original at float32 precision.

When to use this

• You want a Mamba-2 2.7B-class state-space model at reduced storage footprint
• You need reproducible, cryptographically verifiable reconstruction (a deterministic decode to the SHA-256-pinned validated artifact), not opaque "approximately equivalent" PTQ
• You're integrating into your own stack and want a drop-in HF folder after uc unpack

When NOT to use this

• You need raw fp16/bf16 -- fall back to state-spaces/mamba2-2.7b
• You're on hardware without 5-bit unpacking support -- currently CUDA / Apple Metal / x86 with AVX2

License

• This artifact: BUSL-1.1 with Additional Use Grant for research, individuals, and companies under $1M ARR
• Auto-converts to Apache 2.0 four years after release date
• Above $1M ARR running this in commercial production: founder@sipsalabs.com
• Patent posture: two USPTO provisional applications filed April 2026

Citation

@software{sipsa_ultracompress_2026,
  title  = {UltraCompress: near-lossless 5-bit model compression at scale},
  author = {Sipsa Labs, Inc.},
  year   = {2026},
  url    = {https://github.com/sipsalabs/ultracompress},
  note   = {Two USPTO provisional applications filed April 2026}
}

Contact

• Commercial license: founder@sipsalabs.com
• Security: security@sipsalabs.com
• General: founder@sipsalabs.com

---

This artifact is part of the SipsaLabs cross-architecture verified matrix. Browse the full collection on the org page.