Model Support Roadmap

Model Support Roadmap

Last updated: 2026-08-01.

This page mirrors the current repository roadmap. The authoritative status source is docs/MODEL-SUPPORT.json, and current source code takes precedence over this Wiki if they disagree.

Status

Model or path	Status	Current boundary
Deterministic QEMU model-v1 path	Fixture only	Validates OS/runtime contracts. It is not transformer inference or a hardware benchmark.
xaios.model.v2 tooling	Interface only	Package parser/writer and interface tests exist; no production model executes from it.
Qwen 3.6 27B	Interface only	Primary scalar correctness bring-up target; tokenizer, logits and deterministic decode parity remain incomplete.
Kimi K3 text	Interface only	Active target for KDA, Gated MLA, exact top-16 MoE, shared experts and MXFP4; text inference is not available.
Kimi K3 multimodal	Roadmap only	Vision preprocessing/tower/projection and multimodal parity are a separate milestone.
DeepSeek V4 Flash 0731	Roadmap only	Planned adapter target. The exact official release/config/tokenizer sources must be verified first.
GLM 5.2	Roadmap only	Planned adapter target; importer, tokenizer, operators, state and parity work have not started.

Fixture only, Interface only and Roadmap only do not mean model support. A model becomes supported only after its official tokenizer, tensor import, operators, state, logits, deterministic decode and physical-hardware gates pass.

Requested Model Targets

Kimi K3

Kimi K3 is an active first-class architecture target with separate text and multimodal milestones. Exact top-16 target routing remains authoritative; future predictive routing may influence prefetch only. Full K3 support cannot be claimed until both text and multimodal acceptance gates pass.

DeepSeek V4 Flash 0731

The requested label is recorded as a roadmap target. Before implementation, XAIOS must locate and pin the immutable official configuration, tokenizer, tensor index and model revision, then derive a separate architecture adapter from those sources.

GLM 5.2

GLM 5.2 is a roadmap target. Its adapter must be based on a pinned official configuration and must validate the configured sparse-attention, MoE and long-context state requirements rather than assuming a generic transformer.

Qwen 3.6 27B

Qwen 3.6 27B is the primary real-model correctness target. Its immutable official source revision must be pinned before implementation, followed by tokenizer, layer, logits, deterministic decode and physical-hardware gates.

Implementation Order

1. Complete model-v2 import and package-owned tokenizer support.
2. Pass the current small Qwen-family scalar correctness gates.
3. Implement Kimi K3 text, then its separate multimodal milestone.
4. Pin official sources and implement separate adapters for DeepSeek V4 Flash 0731 and GLM 5.2.
5. Add optimized Apple and Intel backends only after scalar parity.
6. Publish physical-hardware results only under the repository benchmark contract.

Sources and Detailed Plans

• Repository implementation roadmap
• Model-v2 specification
• Benchmark contract
• Kimi K3 official repository
• Kimi K3 official model
• GLM 5.2 official model

An immutable official source has not yet been pinned for the exact label DeepSeek V4 Flash 0731. Its inclusion here records intent, not compatibility or implementation.