Nemotron 3.5 Super VL plugs into Deep Agents with LangChain benchmarks

LangChain says NVIDIA's Nemotron 3.5 Super VL, an open-weights model with 120 billion parameters of which 12 billion are active per token, now works as the model inside Deep Agents, LangChain's ready-made agent loop with built-in tools (LangChain post). The connection runs through ChatNVIDIA, LangChain's wrapper for models served from NVIDIA's API catalog or self-hosted in NVIDIA's NIM containers. The integration docs show the whole path: create a ChatNVIDIA model with the ID nvidia/nemotron-3.5-super-vl-120b-a12b, an 8,192-token completion limit and thinking switched on, then pass it to create_deep_agent. Because the model accepts images as well as text, the same agent can be handed an image inside its message, which is where the vision in the announcement comes in.

The numbers in the post come from LangChain's own tests: an 87% average pass rate on its Deep Agents behavioral suite, and, on a benchmark it calls Harbor-lite, 55% less total task time than DeepSeek V4.1 Flash with what it describes as similar success. Neither the post nor the linked docs say what those suites contain, how many tasks were run or what the success rates were, so they describe speed and reliability inside LangChain's harness rather than a comparison a reader can reproduce. The docs also note that their example uses the default Deep Agents harness without a model-specific profile. A profile adjusts the harness for one model, for instance its prompts, tools or middleware; NVIDIA's earlier write-up of one for Nemotron 3 Ultra, covered separately, reported that it raised that model's evaluation score from 94 to 96 out of 127. The post does not say whether its own results used such a profile.