Skip to content

Publication summary: Qwen3.8 Flash Next shortlist trial

This is derivative copy. The dated finding and retained native artifacts are authoritative.

Canonical facts

  • Model identity: RadixArk/Qwen3.8-Flash-Next-NVFP4@7b719225242aacd3dbd3f9407468c2ee9a9d2594.
  • Runtime identity: SGLang 4ccff141dbe992794f9da6c3aa23535b4f72000d; image sha256:8ac9dedd8c3c98bec7b91f4c2c7b697e4f429e336428dd8167140075e295bae7.
  • Local setup: 2x RTX PRO 6000 Blackwell Max-Q, TP2, C4 configured admission, 262,144 configured tokens.
  • Measurement path: direct online cold-start and serial C1 functional/agentic checks.
  • Managed recipe / reproduction: operator-private recipe; public reconstruction in the exact configuration and run plan.
  • Headline result: startup 65.55 GiB/rank without OOM; direct preflight groups 6/6.
  • Capability result: Qwen scouts were 16/18 greedy and 15/18 native sampling; paired GLM was 16/18.
  • Important caveat: native results pass their own 0.75 floor, but fixture/scorer defects block broad planning/coding comparison.
  • Decision: retain GLM; Qwen is unqualified and no-promotion.
  • Canonical evidence: dated finding.
  • Artifact set: manifest and evidence index.

Screenshot alt text

Qwen3.8 Flash Next NVFP4 running TP2 on two RTX PRO 6000 GPUs loaded without OOM and passed six serial C1 direct checks; C4 admission was configured, not measured. Agentic scouts scored 16 of 18 and 15 of 18, below the campaign's required perfect deterministic score; no capacity or throughput result is claimed.

Claim ledger

Public claim Conditions Evidence
Qwen startup and direct preflight passed TP2, cold start, serial C1 checks; C4 configured only finding; status
Agentic comparison is blocked 18 observations per arm; campaign requires 18/18 finding; Qwen greedy; paired GLM
No performance claim follows Capacity and performance were not run finding

Short post

Local Qwen3.8 NVFP4: 6/6 serial checks passed on 2x RTX PRO 6000. Harness defects block a coding comparison; GLM retained. https://fakoli.github.io/anvil-serving/findings/2026-09-22-model-shortlist-validation/

Draft copy only; the canonical site URL becomes available after publication.

Reddit variant

Title: Local Qwen3.8 NVFP4 scout: basic checks pass, agentic comparison remains unresolved

Body:

We screened independent benchmarks before trying the pinned RadixArk NVFP4 checkpoint with SGLang TP2. It loaded at 65.55 GiB of weights per rank and passed six serial C1 preflight groups. The Qwen greedy and paired GLM scouts both recorded 16/18, but fixture-state drift and a planning-scorer false negative prevent a broad coding comparison. Capacity, vision and performance tests remain unrun. The exact GLM baseline is restored. The finding links the native traces, runtime identity and open harness ticket.