Skip to content

0033. Qwen3.6 Palette Flip to B60 Unified Backend

  • Status: accepted
  • Date: 2026-07-11
  • Deciders: lusoris
  • Related: PR #2314

Context

The model-serving palette was consolidating onto the B60 unified backend (llama-swap fronting swappable models on the Arc B60), per the GPU unified backend design. The Qwen3.6 generation — a 3.6-27B dense model, a 35B-A3B MoE, and the 3.5-9B — needed placement on that backend along with the parser configuration and the deployment components to serve them, replacing the prior split-backend Qwen palette.

Decision

Flip the Qwen palette to the B60 unified backend as phase P1: serve 3.6-27B, 35B-A3B, and 3.5-9B through llama-swap children on the B60, register the cauda36 (dense) and cauda36moe (MoE) model prefixes consumed via LiteLLM, wire the parser configuration for the new models, and add the deployment components required for the unified topology.

Alternatives considered

  • Keep the prior Qwen3 palette on split/dedicated backends: retains topology divergence the unified-backend design is retiring.
  • A CUDA-first serving path: rejected by the unified-backend design in favour of the Arc B60 llama-swap layer for these tiers.

Consequences

  • Consumers reference the new cauda36* prefixes through LiteLLM; no consumer wiring changes to a card directly (consistent with ADR 0030).
  • The B60 warm-tier thrash constraint applies: only a limited number of tiers stay resident, so palette placement must respect co-residency limits.
  • Advances the GPU unified backend program documented in .workingdir2/gpu-unified-backend-design.md.

References

  • PR #2314 — feat(serving): Qwen3.6 palette flip (P1) — 3.6-27B/35B-A3B/3.5-9B on the B60 unified backend
  • .workingdir2/gpu-unified-backend-design.md, apps/ai/paperclip/templates/deployment.yaml (backend prefixes)