NemoClaw agents, TensorRT Edge-LLM, Dynamo AICR, MOSS 2-step

No big model launch in my feeds today. What did show up is tooling, mostly from NVIDIA, that makes agents and inference easier to ship, plus a small image adapter worth a note.

NemoClaw open blueprints

NVIDIA published NemoClaw, a set of open blueprints that bundle an agent framework, Nemotron models, NeMo optimization tools and the OpenShell secure runtime. The blueprints currently support three orchestration layers: OpenClaw, Hermes Agent from Nous Research, and LangChain Deep Agents. The idea is that you keep the framework you already use and add model optimization, telemetry and governance around it.

The part I care about is OpenShell, which governs what an agent can reach: files, networks, credentials and tools. NVIDIA says agents stay bound by policy even when memory or retrieved context suggests an action they should not take. That is the right place to put a control, because prompts are a weak boundary.

On cost, LangChain reports that it tuned Deep Agents for Nemotron 3 Ultra and got near-frontier performance at about 10x lower cost per run. The tuned profile is on GitHub. NVIDIA says NemoClaw agents run in the cloud, in the data center, on DGX Spark and on RTX workstations, and can sit alongside frontier agents such as Claude, Codex or Cursor.

TensorRT Edge-LLM on DGX Spark

A companion post walks through TensorRT Edge-LLM on the new 64GB DGX Spark. The runtime is a lightweight C++ stack meant to span development on Spark and deployment on Jetson Thor, DRIVE AGX Thor and IGX Thor. The trtedge CLI downloads a model and builds the TensorRT engines, so a chat with Nemotron-3.5 Lightning is a couple of commands, and an OpenAI-compatible server can run locally for prototyping.

NVIDIA also shows DSpark speculative decoding with three draft tokens, lifting mean generation speed from 73.4 to 126.9 tokens per second on HumanEval prompts with peak memory at 27.9GB.

Dynamo on Kubernetes

The third post covers NVIDIA AI Cluster Runtime, which captures GPU drivers, resource allocation, gang scheduling, node health signals and metrics in reproducible recipes with pinned versions. Dynamo 1.4.2 uses TCP for the request plane and ZMQ for KV-cache events, so both its native router and the Gateway API path can make cache-aware routing choices. The flow is five steps: snapshot the cluster, generate a recipe, render a bundle, deploy with Helmfile or GitOps, then validate before any application goes on top.

MOSS-Image in two steps

On the open-weights side, a community adapter on Hugging Face runs MOSS-Image-SFT in two network evaluations with no CFG, against 50 steps with CFG 3 for the base. It is a LoRA built with a community implementation of Intrinsic Distribution Matching, trained for 1000 iterations in roughly two hours on one H200, and the author notes it deviates from the paper in several ways. The card claims legible English and Chinese text rendering. It is a community result, not a MOSS team release.

What I'd watch next

Whether anyone publishes independent cost-per-run numbers for NemoClaw agents beyond the LangChain figure, and how the 2-step MOSS adapter holds up on prompts outside the author's sample prompts.