LFM3-1.7B vs Nemotron 3.5 Lightning 30B A3B: One Gigabyte Against Twenty-Two

Most model comparisons start with a benchmark table. This one has to start with a memory budget, because LFM3-1.7B and Nemotron 3.5 Lightning 30B A3B do not compete for the same slot in a system. NVIDIA's model, released on August 11, 2026 under OpenMDW-1.1, needs about 22 GB in its NVFP4 build. Liquid AI's, announced on October 27, 2026, needs roughly 1 GB at 4-bit and is designed to run on a phone. For any model we do not host ourselves, routing is the control surface rather than the host. Provider list prices pass through at 0% markup.

Everything downstream of that fact follows from it. Nemotron 3.5 Lightning is the more capable model on every benchmark both vendors published — GPQA Diamond at 75.57 against 63.37, SWE-bench Verified at 52.80 against 43.40, as listed in the evaluation panel on NVIDIA's own Hugging Face card. It also carries a context window up to 1 million tokens against 262,144. If your question is which model is better, the answer is unambiguous and this page could end here.

The question worth asking instead is which model has a job in your deployment. A 22 GB model does not run on a phone, a laptop, or a Jetson in any useful configuration. A 1 GB model does not repair a repository. Those are different constraints, and a comparison that only counts benchmark points will tell you to buy the wrong one.

What actually fits where

Nemotron 3.5 Lightning 30B A3B is a hybrid mixture-of-experts model: 30 billion total parameters with roughly 3 billion active, 128 routed experts with six firing plus one shared expert, and Mamba-2 state-space layers interleaved with a small number of full-attention layers. NVIDIA ships BF16 reference weights and an NVFP4 build at about 22 GB that fits a single GPU, and the model runs on B200, H100, H200, A100, GB200, DGX Spark, RTX 5090, and Jetson hardware.

LFM3-1.7B is dense and small: 1.73 billion parameters, 24 layers of gated short convolutions and sparse attention, 3.46 GB in BF16 and about 1 GB at 4-bit. Liquid's card says it fits on a phone or a laptop, and at that size the claim holds up.

• Total parameters — 1.73B dense for LFM3-1.7B; 30B with about 3B active for Nemotron 3.5 Lightning.

• Deployment memory — about 1 GB at 4-bit against about 22 GB in NVFP4.

• Context — 262,144 tokens against up to 1,000,000, with 256K on a single H100 and 262K on hosted endpoints.

• Modality — text only on both; neither ships a vision or audio path in these checkpoints.

• Languages — sixteen named for LFM3-1.7B; English plus five for Nemotron 3.5 Lightning.

• License — LFM Open License v1.0, free below $10 million in annual revenue, against OpenMDW-1.1, which permits commercial and non-commercial use without a revenue test.

Nemotron's speed argument is not that it is small

It would be easy to read "Lightning" and "3B active" as the same pitch as a tiny dense model. It is not. NVIDIA's speed story rests on three things the LFM3 model does not have: a Mamba-2 hybrid architecture that scales better in long sequences, multi-token prediction with speculative decoding, and dedicated draft models — DSpark for DGX Spark and low-concurrency setups, DFlash as an alternative.

The published measurements back this up. On a DGX Spark, vLLM with DSpark reached about 115.8 tokens per second single-stream and roughly 422 tokens per second aggregate at eight concurrent requests, against about 79.6 tokens per second without speculation. NVIDIA claims up to four times faster generation than similar-sized models and positions the model on the accuracy-speed Pareto frontier.

LFM3-1.7B's speed advantage is different in kind. It is fast because there is almost nothing to run — 1.73 billion dense parameters, a 1.6 GB KV cache at 128K tokens, and no routing overhead. Liquid reports 1.8 to 2.8 times faster decoding than its own predecessors on tested hardware and 3.6 billion tokens per day on a single NVIDIA B300. That is throughput on hardware nobody puts in a phone.

So the two models are fast for opposite reasons: one by engineering a large model to compute less per token, the other by being small enough that the question is moot. On a server, the engineering wins. On a device, the small model is the only option that exists.

The capability gap, stated plainly

NVIDIA's Hugging Face card lists 81.62 on MMLU-Pro, 75.57 on GPQA Diamond, and 52.80 on SWE-bench Verified for Nemotron 3.5 Lightning 30B A3B. Liquid reports 63.37 on GPQA Diamond and 43.40 on SWE-bench Verified for LFM3-1.7B, and does not publish an MMLU-Pro figure for that model at all.

Two of those three are directly comparable and both favour NVIDIA, by 12.2 and 9.4 points. That is a real gap on exactly the tasks where a small dense model is weakest: graduate-level science reasoning and repository-scale code repair. Liquid's own model card concedes the second, stating the model is not recommended for agentic coding or complex multi-step reasoning due to its small size.

A caveat on the exact digits: NVIDIA's published figures for this checkpoint drift slightly between its release materials and the Hugging Face evaluation panel — earlier write-ups cited 75.44 and 51.56 for the same two benchmarks. The direction and rough size of the gap are stable across versions; treat the specific decimals as vendor-reported and version-dependent rather than as a fixed score.

Where LFM3-1.7B has its own numbers, they are agentic rather than knowledge-based: 71.39 on BFCL V4, 80.36 on ToolSandbox, 68.01 on τ³-bench, 38.42 on WebArena-Verified, 96.66 on IFEval. NVIDIA does not publish BFCL or ToolSandbox figures for this checkpoint, so those cannot be lined up. NVIDIA does report 86% accuracy on PinchBench across 10,000 tasks, completed 30% faster than Qwen3.6-35B at similar accuracy — a claim Liquid has no counterpart to. The two vendors simply measured different things, and only the overlap supports a verdict.

Context is the widest real gap

A million tokens against 262,144 is not a marginal difference. Nemotron 3.5 Lightning's native window is four times longer, and even the conservative single-H100 configuration at 256K matches the LFM3 model's maximum. For workloads built around whole-repository context, long-horizon agent traces, or hour-long transcripts, the small model's window is the constraint before its capability is.

Worth noting, though, that a long window and long-context retention are different claims. LFM3-1.7B reports 65.56 on RULER at 128K, which is modest — it can accept 262K tokens but does not hold all of that context equally well. NVIDIA has not published a RULER figure for this checkpoint in the materials reviewed here, so the retention comparison cannot be made. Do not assume the 1M window translates into 1M of usable attention.

Licenses differ more than the benchmarks suggest

This is the one dimension where the small model's vendor made the less convenient choice. Nemotron 3.5 Lightning 30B A3B is released under OpenMDW-1.1, which permits commercial and non-commercial use without a revenue threshold. LFM3-1.7B is under the LFM Open License v1.0, free for commercial use below $10 million in annual revenue, with a paid commercial license required above that line.

For an individual developer or a startup, the difference is theoretical. For a team inside a large company, or one planning to be large, it is a procurement step that Nemotron avoids entirely. Given that LFM3-1.7B's whole pitch is deployment inside your own product, the license is worth reading before the architecture is worth admiring.

Where the 1 GB model is still the answer

None of the above makes LFM3-1.7B a bad model. It makes it a model with a narrow deployment envelope, and inside that envelope nothing about Nemotron 3.5 Lightning is relevant because the larger model cannot be there at all.

If the product ships to a phone, a laptop, an AI PC without a discrete GPU, or a field device, LFM3-1.7B is not the compromise — it is the only member of this pair that runs. A 262,144-token context on a device that fits in a pocket is genuinely unusual, and the sixteen-language vocabulary covers most regional deployments. For tool calling, structured extraction, classification and local assistants, that is enough to build a product.

The right pattern for most teams is both, tiered. Keep LFM3-1.7B on the device for the fast, private, schema-shaped calls, and send the reasoning-heavy or repository-scale work to a server-side model. That is an architectural decision, not a model preference, and it happens to be how most shipping agents are already built.

Where a routing layer fits is the server tier. Nemotron 3.5 Lightning is served through NVIDIA's own stack and a number of third-party platforms, and OrcaRouter's role is the one it plays for any model we do not host ourselves: a single key and a single bill across more than 200 models, provider list prices passed through at 0% markup, and automatic failover across providers so a long-running agent does not fail because one endpoint degraded. Neither LFM3-1.7B nor Nemotron 3.5 Lightning is on our catalogue today — both are open-weight downloads, and the small one is meant to run on hardware you own.

The floor is the point

If you have a GPU and your problem is capability, Nemotron 3.5 Lightning 30B A3B is the better model by a clear margin and there is no argument to have. It reasons better, repairs code better, holds four times the context, and its license has no revenue test.

If you have a device and your problem is that nothing else fits on it, the comparison reverses completely. A 22 GB model is not a candidate, and the relevant numbers become 1 GB, 262,144 tokens and a tool-calling profile strong enough that Liquid published it as the model's headline capability.

The open question is whether Nemotron's efficiency recipe compresses further. A 30B model at 3B active already runs on a Jetson; a genuinely strong 2B-class model in the same family would put these two in the same conversation rather than on either side of a memory wall. Until then, the honest comparison is not between their benchmark scores but between the hardware each one requires — and only one of them requires none.