Mimir remembers.
Every inference runs on your hardware, under keys only you hold. Mimir is agentic AI built on GLM 5.2, inside the enclosure, against your sealed objects, with every action on the audit line. First token in under 800 ms, or none at all.
Intelligence that stays home
No core function calls a cloud. Models run on your GPUs, in your enclosure, air-gapped if you choose. If the hardware isn't yours, the answer isn't either.
The model you evaluated is the model that runs. Weights are content-addressed and sealed; a swapped model is a signature mismatch, not a surprise.
Mimir's agents act through Midgard's permission fabric. Every tool call is scoped, attributed and logged. An agent can do nothing you could not audit.
The well draws from the fabric
Mimir reasons over the object fabric, not over copies exported to someone else's index. Retrieval respects sealing and governance: an agent sees exactly what its operator may see, nothing more.
Answers cite objects, and objects carry lineage. Ask how Mimir knows, and it shows you the provenance line, down to the transform that produced the row.
$ mimir ask "which vessels changed flag twice in 30 days?"
3 objects · vessel-0417, vessel-0562, vessel-1188
provenance: feed-atlantic-raw → forge/tx-2214 → fabric
model: glm-5.2 · sha256:9f31…c04a · local
first token: 612 msOn the record
| Property | Specification |
|---|---|
| Model | GLM 5.2 · content-addressed weights · sealed at rest |
| Inference | Local only. Zero outbound calls. Air-gapped operation supported. |
| Latency target | first token < 800 ms on reference hardware |
| Retrieval | Governed access to the object fabric; answers cite objects with lineage. |
| Audit | Every agent action and tool call on the signed audit line. |
| Competes with | Palantir AIP. |