Queue options with per-queue polling intervals, workflow/step count metrics, and easy ListRegisteredQueues helpers.
Historical fullnode sync, encrypted mempool with transaction filtering + PVSS batch encryption, Multi‑Asset faucet, and Move reflection support land in Aptos.
Big GPU boost: new GPU device API for direct GPU-to-GPU, Direct NIC/DPU data path with CUDA (GDA/GDAKI), plus support for ConnectX-9 and CUDA 13.
Now supports importing modpacks via external sharing, pinning versions, and new Control Layout Editor options like focus toggle and custom text sizes—plus adjustable video background volume.
Now supports Gemini 3.0 Flash/Flash Thinking, adds an AI Studio RTL toggle, model-specific new‑chat shortcuts and a custom hotkey recorder, plus a customizable Windows installer.
Mesh shaders land (Vulkan full, Metal/DX12 passthrough), multiview arrives across major platforms with selective masks, adapter enumeration goes async, and a new LoadOp::DontCare unlocks faster passes.
Big memory wins: memory hot-plugging with virtio-mem, persistent memory via virtio-pmem, and virtio-balloon free page reporting/hinting.
Smart Coercion is now stable and a new RethrowHandlerPlugin lets you bubble errors to your framework.
FP8 Q/KV cache and SM100 FP8 FMHA land, plus TRT‑LLM FMHA v2 and MxInt4×BF16 MoE support, a unified multi‑GPU comms API, and a CUDA‑graphs‑friendly sampler.
Big drop: dynamic per-request agent switching with per-agent LLM/tool limits, new S3-compatible artifact, MongoDB and Milvus storage, Hunyuan model and DeepSeek thinking support, plus a built-in web fetch tool.