Nemotron 3 Ultra by NVIDIA

Big release. What’s interesting to me is less the “bigger context window” headline and more what it means for actual agent runs, where most of the work is planning, tool calls, backtracking, and keeping state over time.I’m curious how you’re seeing people use Nemotron 3 Ultra alongside retrieval or external memory. With a 1M context window, does that layer become less important, or does it just shift toward deciding what should live in memory vs what gets passed straight into the run?

A lot of frontier models are improving raw reasoning, but context management still feels like a separate bottleneck.Have you seen longer-horizon agent workloads benefit more from the model improvements themselves, or from better retrieval and memory layers around them?

550B params (55B active), 1M context, 300 tok/sec. probably the strongest US open-weights model out there right now - and it's currently available for free on @Kilo Codeouch.

NVIDIA just shipped Nemotron 3 Ultra, a 550B open frontier model purpose-built for long-running AI agents.Most frontier reasoning models are optimised for single-turn accuracy. Agentic tasks are different: agents plan, call tools, delegate to sub-agents, handle failures, and pass history back into the model across many turns. As sessions get longer, token costs compound and models start losing the thread.Nemotron 3 Ultra addresses this with a hybrid Mamba-Transformer architecture that handles long-context sequences without losing recall, and NVFP4 quantisation that delivers 5x higher throughput per GPU compared to BF16 on Blackwell.Here's what ships:550B total / 55B active parameters via LatentMoE so you get frontier reasoning without activating the full model on every tokenUp to 1M token context window handles large codebases, long tool-call chains, and multi-document synthesis nativelyMulti-token prediction layers reduces generation time on long outputs and multi-turn workflowsPost-trained for OpenClaw, Hermes Agent, and LangChain Deep Agents accurate across agent harnesses, not just chat benchmarksMulti-Teacher On-Policy Distillation trained with dense feedback from 10+ domain-specific teacher models across code, math, and tool useFully open weights, synthetic training data, and post-training recipes all released under OpenMDW-1.1P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends