Frontier AI, running on hardware you control.
Eighteen months ago, running a model competitive with GPT-4 on your own hardware wasn't realistic for most organisations. That's changed. Open-source AI has caught up with the proprietary models for most professional work — which means there's no longer a quality tradeoff in keeping your data local.
The capability gap closed. The licence advantages didn't.
The leading open-source AI models — DeepSeek, Qwen, Llama — now benchmark within a fraction of the proprietary frontier (Claude, GPT-4) on document analysis, drafting, and reasoning tasks. The quality argument for US cloud AI has largely gone. You can have the same capability on hardware you control.
The licence matters as much as the capability. MIT-licensed models carry no data-sharing obligations, no attribution requirements, no terms that create legal exposure when processing sensitive client data. The weights download to your server. What happens there is entirely within your control.
And because the open-source frontier ships fast — new models every few months — when a meaningfully better model is released, we update your deployment. You don't wait for a cloud provider's approval process. You don't renegotiate a contract. The hardware keeps earning through every generation of the software.
Three tiers. What each one can do.
The right configuration depends on the size of your team, the number of concurrent users, and the complexity of the tasks. Here's what each tier runs, and what it's capable of.
1× NVIDIA L40S
48 GB VRAM · ~$10,000–$12,000 hardware
Runs: Qwen 2.5 32B Instruct at full quality, or Llama 3.3 70B at reduced precision.
Concurrent users: 1–2 comfortably.
Good for: Document Q&A, email and correspondence drafting, meeting summarisation, straightforward research tasks. Strong on single-document work; less suited to complex multi-document reasoning across large corpora.
Best fit: Small firms up to ~10 staff, or a single department pilot before a larger rollout.
2× NVIDIA L40S
96 GB VRAM total · ~$18,000–$22,000 hardware
Runs: DeepSeek V4 Flash — 284 billion total parameters, 13 billion active per query (Mixture of Experts). 1 million token context window. MIT licensed.
Concurrent users: 3–5 comfortably.
Good for: The full professional services application suite — complex multi-document due diligence, long contract analysis, drafting from templates, research synthesis across large precedent banks, billing capture. Handles the tasks where quality genuinely matters.
Best fit: Professional services firms up to ~50 staff. The right starting point for most law firms, accountancies, and healthcare practices.
4× NVIDIA L40S
192 GB VRAM total · ~$36,000–$44,000 hardware
Runs: DeepSeek V4 Flash at full precision across more parallel sessions, or multiple specialised models simultaneously — a different model per department or application type.
Concurrent users: 10+ comfortably.
Good for: Everything the Standard configuration handles, at higher volume. Multiple departments running different applications simultaneously. High-throughput document processing pipelines. Organisations where AI usage is already established and the bottleneck is capacity.
Best fit: Firms of 50+ staff, or organisations with multiple departments and high concurrent demand.
DeepSeek V4 Flash — what it means in practice.
The Standard and High Throughput configurations run DeepSeek V4 Flash — currently the best open-source model for professional services work. The key numbers: 284 billion total parameters with only 13 billion active per query, via a Mixture of Experts architecture. A 1 million token context window — large enough to hold an entire contract bundle, a year of correspondence, or a full matter history in a single session.
In practice, what this means: it can read a 400-page contract bundle and answer specific questions about it. It can draft correspondence in your firm's style after seeing 10 examples. It can compare two versions of a document and flag every material difference. It can synthesise research across dozens of sources and produce a structured summary.
It benchmarks within a fraction of Claude Opus 4.6 on the tasks professional services firms actually care about. The quality is real — not a compromise made to achieve sovereignty.
DeepSeek V4 Flash at a glance
- 284B total parameters, 13B active per query
- Mixture of Experts architecture
- 1 million token context window
- MIT licensed — no data sharing obligations
- Runs on standard NVIDIA GPU hardware
- Served via vLLM, OpenAI-compatible
Your documents are the system's knowledge base.
The AI model knows how to reason and write. What it doesn't know — until we tell it — is your firm's specific documents, templates, precedents, and history. We close that gap with retrieval-augmented generation: your entire document library is indexed so the system can find and use the right documents for each query.
When someone asks the system to draft a letter, it retrieves your firm's relevant templates and recent examples, and drafts from those. When someone asks it to review a clause, it retrieves your precedents for similar clauses and flags deviations. It answers using your actual documents — not a generic training dataset.
From day one
Index your existing document library. The system immediately knows your precedents, your templates, your past matters. Useful from the first day of deployment. New documents added to your library are searchable within minutes — no retraining needed.
After 6–12 months
Once real usage has accumulated enough high-quality examples, we fine-tune the model on your firm's specific patterns — locking in your style, your terminology, your standards at the model level. Each month of use makes the system more specifically yours.