Lossless 5-bit transformer compression for defense and aerospace AI
DoD acquisition, on-platform inference, and SR-11-7 audit all require bit-identical model behavior across deploys. AWQ / GPTQ / EXL3 cannot deliver it. Sipsa proves SHA-256 verifiable reconstruction across 22 architectures — the regulatory-equivalence floor your acquisition reviewer will demand.
The defense / aerospace inference problem
Edge / on-platform / on-orbit AI inference is GPU-memory-bound by definition. Anduril Lattice nodes, Bolt vehicles, Helsing Centaur, NASA HPSC processors — all of them have to run frontier-class models on smaller embedded GPUs with strict power and thermal envelopes. Every byte of weight bandwidth costs Watts.
Sipsa cuts inference memory ~3-4× at sub-1.5% PPL drift. Same model behavior, smaller per-platform footprint. Plus SHA-256 bit-identical reconstruction means the same model gives the same answer in 100K Bolt deployments — auditable for DoD acquisition and reproducible for after-action review.
What Sipsa delivers for defense / aerospace customers
| Need | Sipsa delivery | Compliance hook |
|---|---|---|
| Bit-identical model behavior across platforms | SHA-256 verifiable reconstruction | DoD acquisition audit; SR-11-7 model risk |
| Frontier-scale on-platform inference | 405B fits on single 32 GB consumer GPU | Sovereign / on-prem / air-gapped deploys |
| NASA HPSC / on-orbit fit | Per-Linear streaming compression with verifiable reconstruction | NASA SBIR Topic ENABLE.2.S26B |
| FedRAMP / IL5 / IL6 trajectory | BUSL-1.1 + Additional Use Grant (no cloud dependency); on-prem / air-gapped deploy supported | Defense customers cannot use shared cloud APIs |
| Inference cost per kW recovered | 3-4× lower memory = 3-4× more inference per kW of compute | Per-mission TCO reduction |
NASA SBIR Phase I alignment
Sipsa Labs has a NASA SBIR Phase I draft submission-ready under Topic ENABLE.2.S26B (High Performance Onboard Computing). The proposal angle: per-Layer streaming compression for on-orbit foundation-model inference on the NASA HPSC processor. Defense / aerospace customers who partner with us at Phase 0 may be cited as Letters of Support in the NASA submission and the AFWERX SBIR Phase I draft (also submission-ready), accelerating both timelines.
Verified at scale
22 architectures verified end-to-end, 40 model artifacts at huggingface.co/SipsaLabs, customer-side reproducible:
pip install ultracompress hf download SipsaLabs/hermes-3-405b-uc-v3-bpw5 --local-dir ./hermes-405b uc verify ./hermes-405b # confirms bit-identical reconstruction uc bench ./hermes-405b # measures TTFT / tokens/sec / VRAM
Phase 0 POC for defense / aerospace AI teams ($5K–$25K, 1 week)
We compress one of your production models. Deliver the lossless artifact + SHA-256 manifest + customer-side uc verify dashboard. You confirm bit-identical reconstruction against your bf16 reference. If we miss the spec, you don't pay. Phase 1 commercial license follows if Phase 0 lands. Compatible with on-prem / air-gapped deploys.
FAQ
Is Sipsa export-controlled?
The Sipsa substrate is general-purpose ML compression and not subject to ITAR or EAR controls in its current form. Customers using the substrate for ITAR-controlled or classified workloads handle their own export-control compliance and deployment-environment certifications.
What about CMMC Level 1 / 2 / 3?
Sipsa Labs is preparing CMMC Level 1 self-attestation for defense customer engagements (target Q3 2026). Higher CMMC levels follow with first signed defense contract.
How does Sipsa fit on-orbit / NASA HPSC processors?
Per-Layer streaming compression means the substrate works on radiation-hardened processors with limited DRAM. We have a NASA SBIR Phase I draft submission-ready specifically for this fit (Topic ENABLE.2.S26B). Earliest hardware-in-the-loop validation: Q4 2026 if NASA Phase I awards.