Agents own the news: Nvidia's open agent models, Amazon's TorchNeuron, and the plumbing boom

Across today's AI digests, one story dominates: the infrastructure for agents. Not new chatbots - the models, runtimes, and tools that make agents actually run.

Nvidia and Amazon go to war over the agent stack

Nvidia launched new open models designed to run AI agents in the enterprise and directly on PCs, pushing agentic workloads beyond the data center. Amazon answered yesterday with its biggest shot at Nvidia yet: the TorchNeuron dev tool, a serious challenge to Nvidia's dominance of the AI training and inference stack.

Both stories broke the same day - and both are about the substrate, not the chat interface. Meanwhile the model layer beneath the app builders is being reshuffled: Replit is no longer working with Anthropic, with OpenAI now its default, per the CEO.

Small MoEs built for local agent loops

On HuggingFace, the trend is MoE models with tiny active parameter counts - exactly what local agent loops need. LiquidAI shipped LFM2.5-24B-A2B: a 24B reasoning MoE with only 2B active params, 131K context, and 184 tok/s on an M5 Max, built specifically for local agent loops. Edge0 followed with a 35B/3B-active MoE preview that is already trending.

The counterpoint is DeepSeek-V4.1-Flash, the 763B flagship trending #1 by views - at 200GB+ even in 2-bit, it cannot run on a single box. That gap is precisely why the small-active MoEs matter.

Agent plumbing floods Product Hunt

The Product Hunt board today is almost entirely agent infrastructure. Exa 3.0 raises the bar for agent search with richer data at lower cost. Composio Instant lets AI agents pay for their own tools. Praex is a coding agent that runs on its own models - a real cost, latency, and control bet.

On GitHub, headroom - context-window management at 76.4k stars - sits at #1 trending, n8n is trending on its MCP client/server (the agentic glue between tools), and Kronos, a foundation model trained on the "language of financial markets," keeps vertical foundation models in the conversation. The day's verdict from the trending list: the money is moving from chat wrappers to model-level and agent-plumbing innovation.

Give agents better senses and a wallet - that is today's product thesis, with almost no end-user wrappers in sight.

Signal to watch: whether the small-active MoE wave - 2B–3B active params, 100K+ context - becomes the default substrate for local agents, and whether agent payments (Composio) move from novelty to plumbing.