Candidate package · vLLM XPU · 4× Intel Arc Pro B70

Qwen3.8 Flash-Next FP8 with no speculation, never-routed experts host-placed, both reference Triton kernels restored and the W13 MoE tile at the base width, on four Intel Arc Pro B70 cards

Qwen's 125B-A6B hybrid-attention MoE, served from its official FP8 weights across four Arc Pro B70 cards, with no speculative decoding. The n-gram table and embeddings live in host memory, every hot expert stays on the cards, the experts a routing census never selects are parked in host memory behind a per-expert table in the MoE kernel, and the two Triton kernels the XPU port had replaced are restored to the model's own reference implementations. The record over the previous one is a removal: a per-phase MoE tile that had been adopted bundled with an unrelated change, and never isolated, cost 2.1% once it was measured alone. One line of a tuned configuration file, no source change, outputs bit-identical across three servers.

What we measured Lab-measured

34.5tok/s class-balanced decode median

median of prompt-class medians, 99 inter-token intervals after TTFT, fixed realistic suite run once cold; A326 fresh-server repeat, 2026-09-08. proof file

Model
Qwen3.8 Flash-Next official FP8 export, TP4/EP4, deterministic full-decode graph, no speculation, never-routed experts host-placed, hyper-connection glue and QSA pre-indexer restored to the reference kernels, W13 per-phase MoE tile set to the base width
Publisher
Qwen
Checkpoint
Qwen/Qwen3.8-Flash-Next-FP8
Compression
FP8 block-128 weights / BF16 KV
Software
vLLM XPU
Cards
4× Intel Arc Pro B70 32 GiB
Model weight bytes
Operating systems
Linux
Delivery
native
Good for
general coding deterministic serving research
Published
2026-09-08
Clean-host replay
not yet

Still missing before this becomes an install guide

  • built and replayed container image
  • clean-host beginner install path (the recovery flow now in the guide covers failures during a run, not a first install)
  • container recipe (not published for this line; see CONTAINER-STATUS.md for the shared torch-ABI blocker)
  • decode, prefill, and TTFT context sweep
  • installable dependency hash lock
  • non-originating-host replay
  • record-gate replay scripts (this guide pins the packets and evidence; the MTP1 sibling carries the replay helpers)
  • record-specific model acquisition helper

Measured performance profiles Not published

No qualified structured context or depth profile is published for this package. Diagnostic evidence may still be linked under “What to know” or in the full guide; nothing is estimated in its place.

Many people at once Not published

No qualified multi-user aggregate profile is published for this exact package. Diagnostic or unsupported boundaries may still appear under “What to know” or in the full guide. Nothing is interpolated or promoted from a different model, quantization, runtime, or card count.

How much faster could this get? No projection

This package's measured workload does not map cleanly onto a single model + compression + card-count shape, so no like-for-like projection is shown. The measured numbers above stand on their own.