`uc compress` design spec (v0.2 — target Q3 2026)¶

Status: 🔵 PLANNED for v0.2.

This page describes the planned uc compress command — self-compression of a customer-provided source model. It is NOT in v0.1; this page is the design spec we'll implement against.

uc compress is the most-requested feature in pre-launch design-partner conversations. It's also the feature most-tightly gated on the patent prosecution timeline (we hold the implementation closed until the non-provisional applications are filed in April 2027).

This feature is planned for v0.2 (Q3 2026). To register interest, email founder@sipsalabs.com.

Goals¶

One-command compression of any HF Hub or local-disk transformer model
Reproducibility — every output artifact ships with a complete provenance manifest
Sensible defaults — uc compress <model> Just Works for typical models without per-model tuning
Pluggable compression methods — both Track A and Track B accessible via flags
Resumable — long-running compression jobs can resume after interruption

Synopsis (planned)¶

uc compress <source-model> [--bpw FLOAT] [--method STR] [--track STR]
                            [--output-dir PATH] [--device STR]
                            [--max-batch-size INT] [--seed INT]
                            [--license STR] [--api-key STR]

Options¶

Option	Default	Description
`<source-model>`	required	HF Hub model ID (e.g., `Qwen/Qwen3-1.7B`) or local-disk path
`--bpw FLOAT`	`2.798`	Target bits per weight
`--method STR`	`track-a-row-overlay`	Compression method to apply
`--track STR`	`a`	Which patent-pending track: `a`, `b`, `a+b`
`--output-dir PATH`	`./models/<source-name>-uc<bpw>`	Where to save the compressed artifact
`--device STR`	`cuda:0`	PyTorch device for compression
`--max-batch-size INT`	`8`	Batch size for the calibration data pass
`--seed INT`	`42`	Deterministic seed for reproducibility
`--license STR`	`sipsalabs-research-eval-1.0`	License identifier to write into the manifest
`--api-key STR`	(env: `UC_COMPRESS_API_KEY`)	API key for the remote compression service

What it does (high-level flow)¶

1. Parse the source model: download from HF Hub if needed, or load from local disk
2. Validate the model is a transformer-class architecture (raise otherwise)
3. Compute the FP16 baseline metadata: parameter count, layer breakdown, etc.
4. Submit the compression job to Sipsa Labs' compression service
   (compute-intensive; not feasible on a customer's local laptop in v0.2)
5. Stream the compressed artifact back as it's produced
6. Write the artifact to disk:
     - `model.safetensors` — compressed weights
     - `tokenizer/` — pre-loaded tokenizer (copy from source)
     - `config.json` — pre-loaded config (copy from source)
     - `ultracompress.json` — provenance manifest
     - `LICENSE` — per the --license flag
7. Verify SHA-256 of the produced artifact against the manifest
8. Print a summary

Example¶

# Default settings — compress Qwen3-1.7B to 2.798 bpw via Track A
uc compress Qwen/Qwen3-1.7B

# Output:
# ./models/<source-model-name>-uc<bpw>/
#   ├── model.safetensors           (1.04 GB)
#   ├── tokenizer/                  (2.7 MB, copied from source)
#   ├── config.json                 (4 KB)
#   ├── ultracompress.json          (provenance manifest)
#   └── LICENSE                     (Sipsa Labs Research and Evaluation License)

Track B (architectural compression)¶

Architectural compression is the most aggressive variant; it produces a model with substantially fewer trainable parameters but requires a calibration pass on representative training data.

uc compress Qwen/Qwen3-1.7B --track b --output-dir ./models/qwen3-frr311x \
    --calibration-data ./calibration.jsonl

The calibration data is a JSONL file with prompt/response pairs representative of the customer's deployment workload. The compression service uses this to validate that the architectural-compression preserves quality on the customer's distribution.

Combined Track A + Track B¶

uc compress Qwen/Qwen3-1.7B --track a+b --output-dir ./models/qwen3-uc-combo

Stacks Track B's architectural compression with Track A's quantization. ~26.7× end-to-end with 68% top-10 retention (cohort median).

Why a remote service vs. local¶

Compression at our quality bar requires:

~10-100 GPU-hours per model (depending on size + cohort verification)
Specialized kernels not yet pip-installable
Calibration data from a curated cohort to maintain quality

Running this on a customer's laptop / desktop is not practical in v0.2. We'll lift the local-only requirement as the methods mature, but the v0.2 release uses a remote compression service.

The customer's source model is uploaded to the service over TLS, processed, and the result returned. Customer data never persists on Sipsa Labs's infrastructure beyond the duration of the compression job (typically minutes).

Authentication¶

Sign up at sipsalabs.com to get a UC_COMPRESS_API_KEY
Free tier: 1 compression per month on models < 7B parameters
Paid tiers: per PRICING_CALCULATOR.md

Reproducibility¶

Every compression run is deterministic given the same seed + same source model. The ultracompress.json manifest captures:

Source model SHA-256
Compression method version
Seed
Calibration data SHA-256 (for Track B + combined)
Compute environment fingerprint

A second uc compress run with the same inputs and same seed produces a byte-identical output (within GPU-arithmetic non-determinism bounds; we publish the bound).

Resumability¶

For long-running compression jobs (Track B at 70B+ parameters, expected to take several hours), uc compress supports resume:

# Submit the job
uc compress Qwen/Qwen3-32B --track b --bpw 2.5 --output-dir ./models/qwen3-32b-frr-2p5

# Job interrupted (say, network drop)
# Resume with the same command:
uc compress Qwen/Qwen3-32B --track b --bpw 2.5 --output-dir ./models/qwen3-32b-frr-2p5
# CLI detects the partial state in the output dir, resumes from the last checkpoint

Privacy + IP¶

Customer's source model: stays customer's. We do not retain customer source models beyond the compression job duration. The current subprocessor list is available on request — email legal@sipsalabs.com.
Compressed output artifact: customer owns it; subject to the license written into LICENSE by uc compress (defaults to Research and Evaluation License; commercial customers use a different license value).
Compression methods: Sipsa Labs's IP, patent-pending. Use of uc compress requires acceptance of a Subscription Agreement provided by legal@sipsalabs.com at onboarding.
Calibration data: stays customer's. Used only for the compression job; not retained.

What this command will NOT do (v0.2)¶

Compress arbitrary architectures outside the transformer family (CNN, GNN, etc.)
Compress encoder-only models (T5, BERT) — defer to v0.3
Compress quantization-aware-training models (already-aware models) — defer to v0.3
Run entirely locally without internet — defer; some path forward via --offline flag with pre-shipped kernels

Pricing for `uc compress`¶

Per PRICING_CALCULATOR.md:

Plan	Cost	What you get
Free	$0	1 compression / month, models ≤ 7B params, research license only
Solo	$99/mo	5 compressions / month
Team	$499/mo	50 compressions / month, 5 users
Business	$1,999/mo	200 compressions / month, 15 users, SLA
Enterprise	$5K-$50K/mo	Custom volume, custom users, SLA, audit logs, custom calibration cohorts

For chip vendors and OEMs: separate per-device royalty model in OEM_LICENSING_TERMS.md.

Migration path from v0.1¶

v0.1 doesn't have uc compress. The migration path:

v0.1 → v0.2: uc compress becomes available for customers on a paid tier
Existing v0.1 reference-model artifacts (downloaded via uc pull) keep working in v0.2

Roadmap¶

Feature	Target
Basic `uc compress` with Track A	v0.2 (Q3 2026)
Track B support (architectural compression)	v0.2
Combined Track A + Track B	v0.2
Resumable jobs	v0.2.1
Custom calibration cohorts (enterprise tier)	v0.3
Encoder-only model support (T5, BERT)	v0.3
`--offline` mode (local-only compression)	v1.0+

Open questions¶

API surface for programmatic access: from ultracompress_cli import compress(model_id, bpw, ...) or stay CLI-only? Lean toward CLI-only for v0.2; add Python API in v0.3 if customer demand justifies.
Custom calibration cohorts: how do we expose them safely? Tentatively: customers upload calibration JSONL with stricter NDA/contractual terms, similar to enterprise data-handling.
Performance benchmarks per output: should uc compress automatically run uc bench on the output before declaring success? Yes; small --bench-on-finish flag (default on, suppressible).
Per-customer compression-method versioning: should customers be able to pin to a specific method version (e.g., --method-version 1.2.0)? Yes; expose via flag.

These are open. Customer feedback welcome — file an issue at github.com/sipsalabs/ultracompress or email founder@sipsalabs.com.

Last updated: 2026-04-25 evening. This is a design spec for v0.2; revise as implementation progresses.

uc compress design spec (v0.2 — target Q3 2026)¶