Compressed. Attested. Portable.
Compressed
1.444x lossless: the state is preserved exactly and the model’s output does not change. GPU-native, hardware-accelerated.
Attested
Cryptographic credentials bind every saved state to the model that produced it. Signed audit trail.
Portable
Saved state reloads across your NVIDIA estate, any provider, any framework. No vendor lock-in.
Your model. Your hardware. Your benchmarks.
The math
Lossless by default. No application code changes.
A higher-ratio mode is available for workloads that do not require exactness. It is opt-in, never the default, and we do not recommend it where the output has to be defensible.
50-question contract comprehension · Side-by-side comparison
100% baseline match. Zero differences.
| Config | Accuracy | Compression | Baseline Match |
|---|---|---|---|
| Baseline (uncompressed) | 95.5% | 1.0x | N/A |
| Revyzor | 95.5% | 1.444x lossless | 100% |
50-question legal contract · Qwen 2.5 7B · vLLM 0.19 · every answer compared side-by-side
Validated on vLLM (A100 80GB) and SGLang (A100 80GB) · Also validated: Qwen 3B/14B · Gemma 12B
Needle-in-a-haystack · 64K context
Every needle found. Identical to baseline.
| Method | NIAH Score | 64K Context |
|---|---|---|
| Baseline | All needles | 1x |
| Revyzor | All needles | 3.4x compression |
Lossless fidelity
Not “close enough.” Byte-for-byte identical.
One flag. Full roundtrip.
python -m vllm.entrypoints.openai.api_server \
--model your-model \
--kv-transfer-config '{
"kv_connector": "RevyzorConnector",
"kv_connector_module_path": "revyzor_vllm_connector",
"kv_role": "kv_both"
}'Revyzor integrates via vLLM’s KVConnectorBase_V1 interface: the same documented extension point NVIDIA’s own KVBM (Dynamo), LMCache, and FlexKV use. It composes with cache tiering rather than competing with it: every tier holds losslessly compressed, signed state, and every offload transfer moves ~30% fewer bytes.
Cache Certification
Every saved state is cryptographically certified.
DID
Model identity. Weights, config, adapters: one cryptographic identifier.
SHA-256
KV cache hash. Cryptographically signed: change one byte and the signature fails verification.
Ed25519
Signed credential. Model, org, timestamp, validity; independently checkable.
Revoke
Compromised model? Revoke its DID. All cached states invalidated instantly.
Regulators ask “show us what this AI was working with.” The credential is the answer. Learn more →
Session restore · Measured
Restoring an evicted session costs 56.2 ms.
Qwen 2.5 7B bf16, a 5,097-token contract, 292.3 MB of raw state, local NVMe on an L40S. Ten iterations, first dropped. Every arm asserted bit-identical before it was timed.
| What happens | Mean | Note |
|---|---|---|
| Re-read the document from cold | 416.6 ms | Caching off |
| Evicted session, cache pool flooded | 441.5 ms | The realistic miss |
| Revyzor restore, compressed | 56.2 ms | 33.5 disk + 7.9 H2D + 15.3 decode |
| Same restore, stored uncompressed | 60.2 ms | Compression is a net latency win |
| Still resident in the engine’s prefix cache | 39.5 ms | Faster than us, and free |
The last row is on this page deliberately. If your sessions stay resident in GPU memory, the serving engine already handles them and you do not need us. Revyzor is for state that has been evicted, has to survive a restart, or has to be produced again months later.
See it working in real time.
Two live demos. Two platforms. Same lossless compression. Cosine 1.0000.
Demo 1
vLLM Legal Contract Comprehension
5 questions on a real legal contract. Revyzor compresses KV cache on every query. Model answers through reconstructed cache.
Demo 2
NVIDIA NIM on H100
Revyzor running inside NVIDIA's official inference stack. Llama 3.1 8B in FP8 on H100, answering questions about Revyzor itself.
Early access
Pilot on your workload.
We run Revyzor on your model, your data, your hardware.
You see the compression ratio and accuracy before committing to anything.
Exact lossless compression for AI in regulated environments.
State reloads exactly as captured, across server restarts. Byte-for-byte checked at capture. Cosine similarity 1.0000. Validated across 4 model families, 3 GPU architectures, 2 serving frameworks, 1K to 128K tokens.
Provider agnostic
Deploy anywhere. Save everywhere.
Of the cache freed on every GPU, bit for bit identical. The same hardware carries a longer context, or more sessions at once.
Faster to restore an evicted session from storage than to re-read the document, measured on a 5,097-token contract.
One connector. Every major inference platform.
Cerebrium and InferX live · Others integration-ready
Get early access.
Join the waitlist for Revyzor. We'll notify you when pilot slots open.
Or email directly: sales@revyzor.com