Newly scopeable · 29 July 2026

Kimi K3, evaluated under your custody boundary.

Official weights and serving guidance make a sovereign evaluation possible. They do not prove fit, quality, security, throughput or production readiness for your work.

Official weightsArchitecture planANULUM run pending

01Licence and checkpoint custody02Workload and tool tests03Rack and site feasibility04Stop / proceed evidence

The engagement produces a decision, not a model endorsement.

K3 is one replaceable component inside a complete private-AI system.

Fixed evidence contract

Admission record

Exact checkpoint, source, hashes, licence version, serving stack, permitted use and unresolved obligations.

Workload contract

Representative tasks, context, modalities, tool schemas, concurrency, quality, refusals, leakage checks and stop conditions.

Architecture envelope

Accelerator class, interconnect, storage, network, power, cooling, operations and capacity assumptions with headroom.

Evidence packet

Configuration, raw results, failures, limitations, decision log and a recommendation to stop, narrow, compare or proceed.

Three evaluation tracks

TRACK A

Remote feasibility

Licence, workload, architecture and site-envelope review without client data or a performance claim.

TRACK B

Controlled benchmark

A sanitised or approved corpus on an agreed environment, with frozen runtime and acceptance criteria.

TRACK C

Private acceptance

On client-controlled infrastructure, with client authority, representative material and production-readiness gates.

Hardware boundary: the current public model record lists an official eight-accelerator starting path. Exact memory, topology, context, throughput and facility requirements must be recomputed for the selected checkpoint and runtime.

Differentiator

Replaceability is part of the design.

The retrieval, identity, policy, evidence, backup and exit layers should not become inseparable from K3. The same acceptance set can compare K3 with a smaller or differently licensed model.

no silent checkpoint or engine change
no benchmark borrowed from a different system
no agent authority without explicit tools and human gates
no production claim before site-specific acceptance

Primary public sources

Sources were reviewed on 29 July 2026. Publisher claims remain publisher evidence until independently measured by ANULUM on the declared system.