SGLang added serving recipes for DeepSeek-V4-Flash-Vision and DeepSeek-V4-Flash-0731 on 2×DGX Spark. Flash-Vision has native vision; Flash-0731 supports fp4 and NVIDIA official nvfp4; both run on the b12x kernel library, with docs in the SGLang cookbook.
NVIDIA RTX Spark announced local-agent inference optimizations: up to 1.9× higher llama.cpp throughput on GeForce RTX 5090, 1.2× vLLM on RTX PRO 6000 Blackwell, and up to 1.4× on a two-system DGX Spark cluster—another step for faster on-device and local agent serving.
SGLang says it is working with @zai_org and partners to harden production serving for the GLM-5.3 family. After community and Cloudflare reports, chat-template tool-result reordering now early-exits instead of scanning every block; GLM-5.3 and GLM-5.3-Flash templates are updated and live—material for tool-calling coding-agent stacks.
SGLang announced day-0 serving for AntLing’s open Ling-3.0-flash-Fin (124B total / 5.1B active), a finance-tuned MoE for research, retrieval, valuation modeling, and long-document/spreadsheet report prep, with competitive FinFIRST, FinanceAgent, APEX-Agents, SpreadsheetBench, and τ³-Banking results.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.