SGLang ships day-0 recipes for Qwen3.8-Flash: 176B params, 6B active, verified on RTX PRO 6000 and DGX Spark

NVIDIAAI · x · 2026-09-08

Alibaba's Qwen team launched Qwen3.8-Flash (a preview of the Qwen4 architecture), and SGLang shipped day-0 deployment recipes, each verified on real hardware:

• Architecture: 176B parameters with 51B of N-gram embeddings that scale capacity at almost no extra per-token compute (6B activated per token); the embedding table sits in host memory with async prefetch instead of GPU memory
• Attention: GDN + QSA hybrid attention balances memory efficiency and precise retrieval on long-horizon tasks; Gated Residual gives 4 lanes for inter-layer information flow; trained with Muon
• Verified recipes: 1x RTX PRO 6000 (PLE table in system RAM), 1x DGX Spark (local NVMe), 2x DGX Spark with TP=2, plus NVFP4 support via RadixArk and NVIDIA ModelOpt exports
• Caveat: model support isn't in a tagged release yet — build from PR #36497; the team has fixed several issues reported by the community

Original post →

More from Infra

• AI Infra Meme: While Countries Need TW-Scale Power for Datacenters, People Argue About Lightbulbs — mgostIH · 2026-10-27
• Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
• Cloud release velocity scores: AWS 95, Azure 90, Google Cloud 82 — DavidLinthicum · 2026-10-27
• Selling a 5090 to scale local inference: one business weighs 5 hardware paths — Anarchaotic · 2026-10-27
• AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
• Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27

Infra channel →