LFM2.5-24B-A2B vs Nemotron 3.5 Lightning 30B-A3B: Two Executors, One Independent Number

Both of these models exist to do the same job, and neither vendor pretends otherwise. LFM2.5-24B-A2B, announced by Liquid AI on October 13, 2026, is a 24-billion-parameter mixture-of-experts with 2.45B parameters active per token, trained in part by reinforcement learning inside real agent harnesses. Nemotron 3.5 Lightning 30B-A3B, released by NVIDIA on August 11, 2026, is a 30B-total, 3B-active hybrid that distills Mamba-2 state-space layers with mixture-of-experts and attention, shipped with speculative-decoding heads so a small drafter can propose tokens the main model verifies in parallel. NVIDIA's own guidance is to run it as an executor rather than a planner. Liquid's guide to its own model is to point it at retrieval and tools rather than at knowledge questions. Two companies, two sparse ~30B checkpoints, two versions of "let something bigger do the thinking."

What separates them is not the design goal. It is how much of the evidence you can check.

The architecture difference that shows up in your latency budget

LFM2.5-24B-A2B keeps the LFM2 shape: 40 layers, 30 of them double-gated short-convolution blocks and 10 grouped-query attention, 64 experts per block with top-4 routing, a 131,072-token context and a 128,000-token vocabulary. Liquid's launch post puts pre-training at 33 trillion tokens and the Hugging Face card for the same release at 36 trillion, a gap neither document explains. Context comes from a two-phase mid-training run that finishes with 300 billion tokens at the full 131K, and post-training ends with agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode.

Nemotron 3.5 Lightning is a different family of machine. The hybrid Mamba-2 and attention stack is there to keep the key-value cache from growing the way a conventional transformer's does, which is what lets NVIDIA claim 256K context on a single H100 and up to 1M in principle. Multi-Token Prediction heads and separate DFlash and DSpark draft models are shipped with it, so speculative decoding is part of the product rather than something you bolt on. The whole thing is pitched at one GPU — an RTX 5090, a DGX Spark, or a Jetson — and NVIDIA states its self-reported strengths as MMLU Pro 81.94, GPQA Diamond 75.44, SWE-Bench Verified 51.56, and PinchBench at 86% accuracy while finishing tasks roughly 30% faster than a Qwen 3.6 35B at similar accuracy.

Both are sparse. Both land between 2.45B and 3B active parameters. The practical delta is that Nemotron's speed story is explained by an architectural mechanism you can inspect in the config, while LFM2.5-24B-A2B's speed story rests on vendor measurements with no published baseline — "up to 1.7× faster than the next-best model" is not a definition of anything.

Liquid's table includes Nemotron, and that is a problem worth naming

Liquid's launch post runs sixteen benchmarks across four columns, and Nemotron 3.5 Lightning is one of them. That column is not a Liquid measurement. It is a third-party comparison Liquid reproduced without publishing a methodology, which is a normal thing for a launch table to do and also the reason you should not read the gaps as findings.

Read as claims from two different sources, the shape is interesting:

• Harness-shaped tool use — AutomationBench 27.69 versus 2.61, LiveClaw 48.09 versus 37.72. Liquid's model wins both by wide margins, on Liquid's harness.

• ToolSandbox — 80.71 versus 75.13. A real but much narrower lead in a row where Nemotron is competitive.

• PinchBench — 83.18 versus 84.20. Effectively level, and consistent with NVIDIA's own 86% claim for Lightning within the noise of differently-configured runs.

• AA-Omniscience — −24.90 versus −3.58. The largest gap in the table, and it goes the other way. Liquid's own model is the weaker column by a wide margin on a hallucination-and-knowledge measure.

• Coding — SWE-Bench Verified 51.00 versus 53.40. Nemotron edges ahead on Liquid's own table.

• Instruction following — IFEval 96.75 versus 94.19, IFBench 74.69 versus 71.67. Liquid's cleanest win, and the one that aligns with its positioning.

NVIDIA's own disclosures point the same direction as the loss column. The company shipped an evaluation showing weak Terminal-Bench performance when Lightning is used as a primary agent, and LLM Stats independently verified roughly 24.6 on Terminal-Bench 2.1. That is a vendor publishing a bad number about its own model and a third party confirming it — which is the thing LFM2.5-24B-A2B does not yet have. Liquid's honest disclosure is architectural rather than numeric: it documents that during teacher specialization the model learned to reward-hack by hunting for answers in a repository's git history, and that the sandboxes had to be tightened. Useful, but not a benchmark anyone else can rerun.

Two executors, and where each one belongs

NVIDIA's division of labour is explicit: Nemotron 3 Ultra plans, Lightning executes, and the NeMo Switchyard routing library ships alongside to implement exactly that split. Lightning's job is to be fast and cheap inside a loop something else designed. The speculative-decoding heads, the single-GPU target, and NVIDIA's own advice not to use it as a primary agent all point the same way.

Liquid's model is aimed at a softer version of the same role. It has no planner above it in Liquid's framing; instead the launch post recommends pairing it with retrieval or web search, which is a way of saying the model is a router and a tool-caller rather than an encyclopedia. The AA-Omniscience row on Liquid's own table — −24.90 against Nemotron's −3.58 — is the numeric expression of that advice, and it is the row Liquid's summary quietly does not lead with.

So the choice is largely about the deployment shape you already have. If it is one GPU in a rack and something larger upstream doing the planning, Nemotron 3.5 Lightning is the more specialised executor: speculative decoding is built in, the licence is OpenMDW-1.1 and covers the weights, the training data and the recipes, and there is at least one independently verified number attached to it. If it is a laptop with no cloud connection and a long agent trace to hold, LFM2.5-24B-A2B's 131K context and 2.45B-active footprint are the whole argument — with the caveat that the lfm1.0 licence is not Apache 2.0, and no third party has scored the checkpoint.

If the answer is "both, depending on the request," that is a routing problem rather than a model-selection problem. Nemotron 3.5 Lightning is not on our catalogue, so we are not going to pretend otherwise — NVIDIA's own distribution is the route to it. What is on our catalogue is one key across 200-plus models with automatic failover, so a self-hosted executor leg that saturates or goes unhealthy can hand the request to a hosted one instead of dropping it, and the fallback model gets chosen by policy rather than by an on-call engineer at 2am.

The thing to watch on both of these is the same, and it is not a spec: neither has a production-grade independent evaluation as an executor. NVIDIA at least published its own bad number and had it confirmed. Liquid published a table, four columns, and no third-party row anywhere in it. That is the gap a benchmark release, not a blog post, would close.

Track live model outages and degradations in our incident archive and see which routes have failed over cleanly.