Skip to content

Track DeepSeek V4 Flash 0731 support #16

Description

@Pummelchen

Objective

Validate and implement DeepSeek V4 Flash 0731 through a dedicated architecture adapter.

Delivery dependency

This workstream is queued behind XAIOS and Qwen 3.6 27B. It is additionally blocked on verifying the exact official release identity and immutable source revision.

Current verified boundary

This target is roadmap-only. The exact official release label and immutable source revision have not yet been verified, so implementation contracts must not be guessed.

Progress gates

  • Verify the exact official release name and authoritative source repository
  • Pin immutable configuration, tokenizer, tensor-index, and model implementation revisions
  • Record official architecture identifiers and validate configuration fields
  • Define semantic tensor roles, execution plans, session state, routing, and quantization requirements
  • Implement streaming import and a separate architecture adapter
  • Pass tokenizer, tensor, operator/layer, prefill-logit, deterministic-decode, and session-state parity
  • Pass physical-hardware correctness and benchmark-evidence gates

Blocker

Work beyond source verification is blocked until the exact official model identity and revision are established.

Status source

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

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: cpu-ai-runtimeCPU-only AI runtime, tokenizer, inference path, AI kernels.blockedBlocked by dependency, decision, access, or external state.needs: designNeeds design decision or scope clarification.priority: P2-mediumMedium-priority planned work.type: researchResearch, design investigation, or feasibility work.

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions