The British foundation model
A British foundation model, built from the ground up. Pre-trained from scratch on UK compute, on data we chose and can account for, by the team that runs the governed KV cache. The model and the memory it thinks with are designed together.
Why KVern
Adapt someone else's model and you inherit their data, their architecture and their licence. Every open model we serve was also trained without knowing how it would be run, leaving its KV cache, the state it carries from one token to the next, for someone else to make fast, private and accountable. We have spent our time on that cache. KVern is the model we built knowing it, from the first token of data up.
01
Not an adapted import. KVern's data, tokeniser, architecture and every one of its weights are ours, trained on UK compute by a UK company. No foreign model sits underneath it, and no one else's licence decides how it may be used.
02
A model's state is its KV cache, and on long-context work it is most of the memory you pay for. Building from scratch let us choose an architecture for a small cache per token, and train it to hold quality when that cache is compressed and tiered the way we serve it.
03
Because we built the corpus, we can say where KVern's knowledge came from. It ships with the evidence regulated buyers ask for: its sources, how it was evaluated, and what happens to your context once it has answered.
At a glance
A foundation model, not a fine-tune: every layer, from the tokeniser to the weights, is our own. Parameter count and context length are published in the model card.
How KVern is built
Building from scratch is slower and costs more than adapting an open model. It is also the only way to answer, source by source, where a model's knowledge came from, and to design its architecture for how it will be served rather than inherit someone else's choices.
01Corpus and source register
A pre-training corpus assembled from sources we can account for: UK legislation, the parliamentary record and public-sector guidance under the Open Government Licence, licensed sector material, and filtered open web and code. Every source is logged with its licence, date and filters.
02Tokeniser
A tokeniser trained on the corpus itself, so British spelling, UK legal citations and public-sector vocabulary cost fewer tokens. Fewer tokens is less compute and a smaller KV cache on every request.
03Architecture and scaling experiments
Families of small proxy models fix the design before the main run: attention variant, depth against width, and the data mix. Each candidate is scored on loss and on KV cache bytes per token, measured on our own serving stack.
04Pre-training from scratch
The main run, from random initialisation, on UK compute. Checkpoints are evaluated throughout against the same suites the release is reported on, so regressions are caught during the run rather than after it.
05Cache-aware long-context extension
Context extended with the KV cache quantised, evicted and tiered as it is in production, so the model learns to keep its quality under the conditions it will actually be served in.
06Post-training
Instruction tuning on drafting, summarising, extraction and review tasks from regulated work, then preference tuning for answers that cite their source, say when they do not know, and refuse what they should.
07Release
KVern in the model catalog beside the open models we serve today: on our London fleet, in your own cloud account, or installed on your own hardware, including air-gapped networks.
Evaluations
Leaderboards test models in conditions nobody deploys them in. KVern is reported against four evaluations of our own, each named for the problem it measures, and run on the same governed stack our customers use.
UK law and regulation
Questions on UK legislation, regulator guidance and public-sector policy. An answer only scores if it cites the section that supports it.
Results publish with the benchmark report
Long-context recall
Single and multi-hop retrieval across the full context window, scored with the KV cache quantised and tiered exactly as it is in production.
Results publish with the benchmark report
Containment
Prompt injection, system prompt extraction and cross-session leakage attempts against a governed deployment, scored on what gets out.
Results publish with the benchmark report
Economics
GPU-hours per correct answer at production concurrency, counting failures and retries, against open and frontier models on the same tasks.
Results publish with the benchmark report
Every result ships with its method and harness, so it can be checked rather than taken on trust.
Built for the perimeter
KVern runs wherever kvrun runs. Use it on our fleet, in your own tenancy, or on hardware that never touches the internet. The governed KV cache, the audit log and the compliance record are identical in every one.
See where your workloads run for who pays, who holds the data and what changes in each.
Questions
