A top law firm is reportedly cutting off Harvey and its OpenAI/Claude API connections, moving to its own GPU infrastructure and running Kimi and DeepSeek locally.
If true, that’s a significant moment for legal AI.
But there’s something missing from this conversation.
A local LLM is not a litigation intelligence system
Installing Kimi or DeepSeek on a local server doesn’t suddenly make it a litigation intelligence system. You’ve addressed where the data goes. What about:
- Understanding an entire matter?
- Reconstructing years of events across thousands of documents?
- Connecting evidence to claims?
- Distinguishing what the record actually proves from what a model thinks it proves?
- Verifying every legal citation against real authority?
Those problems don’t disappear because the model is running on your own GPUs.
Offline legal AI built for litigation
This is precisely the distinction we built HIGHCOURT around.
HIGHCOURT runs entirely offline on firm-owned hardware, but it isn’t just a locally hosted language model. The system incorporates legal ontology, matter-wide evidence analysis, litigation reasoning, and offline citation verification. The model is one piece of a much larger architecture.
Owned, not rented
We also chose perpetual licensing rather than per-seat subscriptions because we believe firms should own this infrastructure, not continually rent access to it.
On-premise AI is step one
More firms will move toward local AI. That part makes sense. There’s a big difference, though, between owning the machine running a model and owning a complete system built for litigation.
Getting AI off the cloud is step one. Making it trustworthy for legal work is the harder part.