Track Kimi K3 support

[Pummelchen]

Objective

Implement Kimi K3 as a first-class architecture adapter, with text and multimodal acceptance kept as separate gates.

Delivery dependency

This workstream is backlog work. It starts only after the XAIOS platform milestone and the Qwen 3.6 27B correctness milestone are complete, unless the maintainer explicitly reprioritizes it.

Current verified boundary

Kimi K3 text is interface-only and multimodal support is roadmap-only. No K3 checkpoint currently executes in XAIOS.

Progress gates

• Pin immutable official configuration, tokenizer, tensor metadata, and model implementation revisions
• Implement package-owned tokenizer/import and the separate kimi_k3 architecture adapter
• Implement KDA, Gated MLA, AttnRes, exact top-16 Stable LatentMoE routing, shared experts, SiTU, native MXFP4, and model-specific state
• Pass miniature K3 KDA/MLA/router/expert/reduction golden tests
• Pass real-checkpoint tokenizer, operator, target-token, session-state, and physical-hardware text gates
• Implement expert-addressable storage, cache/prefetch, and exact speculative expert-union tracking
• Separately implement and validate vision preprocessing/tower, projection, multimodal positions, chat-template behavior, and golden image cases

Exactness constraint

Approximate routing must remain opt-in. Authoritative top-16 routing and target-model verification are required by default.

Status source

Execution status is tracked in the XAIOS GitHub Project. Support claims remain governed by docs/MODEL-SUPPORT.json.