Granite 4.2 30B vs Nemotron 3.5 Lightning: the thinker and the doer

Granite 4.2 30B and Nemotron 3.5 Lightning are two 30B open-weight models that shipped fourteen days apart with almost opposite jobs. Nemotron 3.5 Lightning, released by NVIDIA on August 11, 2026, is a 30B-total / 3B-active hybrid Mixture-of-Experts — Mamba-2 state-space layers, MoE, and attention — built to be the execution layer of agent systems: fast tool calls, high-volume loops, subagent work, running on a single GPU. Granite 4.2 30B, which IBM published quietly on Hugging Face on August 25, 2026, is a dense 29-billion-parameter reasoning model — a thinker, with selectable thinking modes and agentic reinforcement learning, also self-hostable. Both are open (Apache 2.0 against OpenMDW-1.1), both fit on hardware a team can actually own, and both are August releases whose benchmarks are still essentially vendor-reported. The interesting question is not which is better but which one you need — a planner, an executor, or both.

That planner/executor split is not an editorial invention; it is NVIDIA's own design. Nemotron 3.5 Lightning was distilled from the far larger Nemotron 3 Ultra specifically so that a frontier model could handle planning while Lightning runs the loop, and NVIDIA shipped it alongside NeMo Switchyard, a routing library that implements exactly that division of labor. Granite 4.2 30B, by contrast, is trying to be the whole agent in one 30B package.

Granite 4.2 30B, the thinker

IBM's flagship Granite 4.2 model is a dense decoder-only Transformer with three reasoning modes — full thinking (wrapped in think>.../think> tags), low-effort thinking, and no thinking — plus OpenAI-compatible tool calling and agentic reinforcement learning that trained it to operate terminals, execute code, and search the web inside a sandbox. Its 128K native context extends to 512K. IBM's card claims 89.17 on AIME 2025, 77.60 on MMLU-Pro, and 57.00 on SWE-Bench Verified (in-harness) — all vendor-reported, none independently reproduced, and no Artificial Analysis entry exists yet. It is, in short, a generalist reasoning model that was pointed hard at agentic work and ships in BF16 plus FP8, NVFP4/MXFP4, and GGUF quantizations so teams can actually run it.

Nemotron 3.5 Lightning, the doer

NVIDIA's entry is a different shape entirely. The "30B A3B" name means 30 billion total parameters with only about 3 billion active per token — a hybrid architecture mixing Mamba-2, MoE, and attention that lets it generate fast and cheap. It ships with speculative decoding built in: Multi-Token Prediction heads, plus DFlash and DSpark draft models, so a small drafter proposes tokens the main model verifies in parallel. NVIDIA claims up to 4x output speed versus similar-size models, and the whole thing runs on a single GPU — an RTX 5090, DGX Spark, or Jetson — or at 256K context on one H100, with the model supporting up to 1M context in principle. Its license, OpenMDW-1.1, covers the weights, the training data, and the recipes.

NVIDIA's own positioning is blunt about what Lightning is not: an evaluation it shipped showed weak Terminal-Bench performance as a primary agent — LLM Stats independently verified roughly 24.6 on Terminal-Bench 2.1 — and the guidance is to use it as an executor, not a planner. Its self-reported strengths are MMLU Pro 81.94, GPQA Diamond 75.44, SWE-Bench Verified 51.56, and PinchBench at 86% accuracy while finishing tasks about 30% faster than a Qwen 3.6 35B at similar accuracy. Every one of those is NVIDIA's own number.

The scoreboard, honestly labeled

The useful comparison is the column shape, not the row values:

• Granite 4.2 30B — ~29B dense, all parameters active; Apache 2.0; 128K native / 512K extended; reasoning modes you can select per request; vendor-reported SWE-Bench Verified 57.00, Terminal-Bench 2.1 29.24; the thinker.

• Nemotron 3.5 Lightning — 30B total / 3B active hybrid MoE; OpenMDW-1.1; 1M context (256K practical on one H100); speculative decoding (MTP/DFlash/DSpark); vendor-reported SWE-Bench Verified 51.56, ~24.6 Terminal-Bench 2.1 (LLM Stats verified); the doer.

Read those rows the way NVIDIA intends: Lightning's Terminal-Bench weakness is a feature of its design — it is not supposed to plan, it is supposed to execute. Granite's numbers are a single model claiming to do both. Neither set is independently reproduced at release; both will be within weeks.

The stack, not the model

The two models are less competitors than adjacent components of the same system. An agent needs a planner to decompose a task and an executor to grind through a thousand tool calls. Granite 4.2 30B is trying to collapse both into one dense model; Nemotron 3.5 Lightning is explicitly half of a two-model stack, with the frontier model as the other half. NVIDIA calls that stack NeMo Switchyard; a routing layer does the same thing with any mix of hosted and self-hosted models.

This is where the honest routing note belongs. Neither model is on OrcaRouter — we route neither Granite 4.2 30B nor Nemotron 3.5 Lightning, and the availability line for both is your own hardware or the vendor's own channels. But the two-model pattern they point to is exactly what a routing DSL is for: compose a self-hosted Granite endpoint for the reasoning passes and a self-hosted Lightning endpoint for the execution loops, put a hosted frontier model in front for the planning that needs more headroom, and drive all three from one API key with automatic failover. You do not need to choose between thinker and doer — you can assemble a stack that uses each where it is strong, and let the router arbitrate.

Which one first

If you want a single model to own, one codebase, one serving stack, and are willing to give it all the reasoning and execution work, start with Granite 4.2 30B — but wait for the first independent benchmark before trusting the 57.00 SWE-Bench number. If you are building a high-volume agent where most calls are short tool steps and latency is the budget, start with Nemotron 3.5 Lightning and let a frontier model or a Granite endpoint do the thinking. If you are building anything serious, the answer is both — which is a decision a 30B-only comparison would never have surfaced.