GENTIAN — MULTI-GPU INFERENCE AND TRAINING SOFTWARE

Longer context.
Cached retrieval.
A record for every answer.

Gentian extends inference and training on multi-GPU production servers, single-node and multi-node — demonstrated on vLLM and Megatron, and not specific to either. It runs on the NVIDIA hardware you already own — no replacement, special mode, or re-platforming.

What Gentian does Verify a record yourself

Three things Gentian does

Gentian is one architecture in two forms: Gentian Software, running today on the NVIDIA servers you already own, and Gentian Silicon, the same design in physical implementation.

Context

Two and a half times the context on a single eight-GPU server — over 1.3 million tokens, where the usual approach runs out of memory. A measured capacity floor, not a speed claim.

Retrieval cache

Ordinary caching collapses as a corpus grows — from roughly 13% hit rate to under 3%. Gentian stays above 95% across the same range, so the cache keeps working as the corpus grows. Hit-rate stability, not throughput.

Records

Every answer can carry a signed record: an enrolled NVIDIA device answered a challenge tied to that exact response, before the signature was made. Anyone can check it offline, years later, after the GPU and the key are gone. It shows the device was present when the answer was authorised.

Check the last one yourself, in two commands

You do not have to take our word for any of it. The public repository holds a real signed record of an AI response and a tampered copy of the same record. The first passes. The second refuses, and names the reason. It runs on Python 3 and cryptography — there is no cybiont software in the verification path at all.

device-presence-bound-v1$ python3 standalone_verify.py valid
{"terminal_status":"PASSED","device_presence_profile":"DEVICE_PRESENCE_BOUND",
 "workload_execution_claim":false,"device_locality_claim":false,
 "performance_claim":false,"vendor_code_used":false, …}

device-presence-bound-v1$ python3 standalone_verify.py tampered
{"refusal_code":"PAYLOAD_DIGEST_MISMATCH","terminal_status":"REFUSED"}

Real output, abridged — the tool prints one line of JSON. Note that its limits arrive as fields, not as footnotes from us.

Get the records and run it → · Two commands, no account, no form.

Product and engineering programmes

Gentian is the product. Governance and Nidus are separate engineering programmes under the same Swiss GmbH. Gentian's silicon work is a separate, longer-horizon research track — not the software offer.

Programme · I

Gentian

Multi-GPU inference and training software

Extends inference and training on existing NVIDIA hardware, demonstrated on vLLM and Megatron. It adds longer-context capacity, stable cached retrieval, and a presence-bound record for each answer. Public measurements and their limits are stated on the product page.

See the Gentian product →

Programme · II

Governance

Controls for AI use in regulated organisations

An independently checkable signed record of a real AI response, published with a tampered twin that refuses and names why — both run offline in two commands. Browser-native DLP, policy enforcement, and governed AI workflows are research under the same programme.

Review Governance stack →

Programme · III

Nidus

Inspectable reasoning for AI-assisted engineering

Governance runtime that mechanises the V-model. Requirements, architecture, traces, proofs, and evidence live in a single decidable artefact verified on every mutation. Public preprint on arXiv (2604.05080).

Explore Nidus →