XGrammar-2 ships strict tool calling for agents, adopted by xAI, DeepSeek and vLLM
vllm_project · x · 2026-09-22
MLC releases XGrammar-2, built for agent workloads
- Introduces Structural Tag, a composable JSON protocol that uniformly expresses OpenAI harmony format, tool calling, reasoning channels and custom output structures, exposed via serving engine APIs
- Efficiency optimizations (cross-grammar caching, repetition-state compression, batching and speculative decoding) keep overhead low even for huge structures
- Adopted by xAI, Databricks, DeepSeek and others; integrated into SGLang, vLLM, TensorRT-LLM and MLC-LLM for strict tool calling
- The vLLM team credits the collaboration for reliable, performant tool calling when serving DeepSeek V4.1
More from coding & agent
- Allie Miller: companies overinvest in AI productivity, ignore workflow handoffs — alliekmiller · 2026-09-22
- OpenAI's Logan Kilpatrick: AI product teams should spend >25% of time on benchmarks — OfficialLoganK · 2026-09-22
- Dev swaps in-game 3D models with Scenario's MCP right from his harness — AIandDesign · 2026-09-22
- Cua AI releases Cua-Bench-S1 benchmark and Cua-S1-Nano/4B computer-use models — ycombinator · 2026-09-22
- Multi-agent scaling may dominate next, calls for OpenAI to publish curves — 1a3orn · 2026-09-22
- GPT-6 prompt cache survives reasoning-effort changes; Codex adapts effort mid-CoT — daniel_mac8 · 2026-09-22