LFM3-1.7B vs Gemma 4 12B: When the Seven-Times-Smaller Model Wins a Coding Benchmark

LFM3-1.7B is a 1.73-billion-parameter text model that Liquid AI published to Hugging Face on October 26, 2026 and announced the following day as part of its LFM3 family launch. Gemma 4 12B is an 11.95-billion-parameter dense model that Google DeepMind released on June 3, 2026, and it reads images and audio as well as text. On the one coding benchmark both vendors publish — LiveCodeBench v6 — the smaller model scores 73.46 and the larger one scores 72.0. Read that twice, because it is the entire reason this page exists and also the easiest number on the internet to misread. Routing does not add a margin to a provider's list price. 0% markup See the pass-through pricing on our pricing page.

Both figures are vendor-reported. Neither has been reproduced by an independent evaluator. They were produced by different teams, on different harnesses, months apart, and Liquid has published no methodology note explaining how its LiveCodeBench run was configured. So the honest version of the sentence above is not "a 1.7B model beats a 12B model at code." It is: two vendors report near-identical scores on a benchmark whose name they share, and the nine-point gap on GPQA Diamond in the other direction suggests the two models are not close to equivalent at all.

What is genuinely comparable, and what is not

Most "X vs Y" pages put two vendor tables side by side and call it a comparison. That only works if the tables overlap. Between Liquid's LFM3 launch post and Google's Gemma 4 model card, exactly two benchmark names appear on both sides with compatible definitions: LiveCodeBench v6 and GPQA Diamond. Everything else is reported by one vendor and not the other, and a blank cell is not a zero.

On those two shared axes the result is a clean split. LFM3-1.7B takes LiveCodeBench v6 at 73.46 against 72.0. Gemma 4 12B takes GPQA Diamond at 78.8 against 63.37 — a 15.4-point margin, which is not noise on an evaluation with a few hundred questions.

That pattern is worth naming rather than averaging away. LiveCodeBench measures a model completing competitive-programming problems, which is a narrow, pattern-heavy skill that a small model trained on the right data can get good at. GPQA Diamond measures graduate-level science reasoning that depends on breadth of knowledge — exactly what parameter count buys you. A 1.7B model can be competitive on the first and will lose the second, and that is a coherent result rather than a contradiction.

The vendor undercuts its own best number

Here is the part that should change how you read the 73.46. Scroll to the bottom of the LFM3-1.7B model card and Liquid writes, in plain language, that the model is designed for local agents, tool calling and structured outputs, and that it is not recommended for agentic coding or complex multi-step reasoning due to its small size.

That is the vendor, not a rival, telling you the ceiling. It sits awkwardly beside a LiveCodeBench score that edges out an 11.95B generalist, and the reconciliation is that these measure different things. Solving an isolated competitive-programming problem is not the same task as holding a repository in context and making twenty coordinated edits. A model can be strong on the first and explicitly unsuited to the second, and Liquid is saying so rather than letting the benchmark imply otherwise.

Take the vendor at its word. If your workload is agentic coding, this comparison should not end with LFM3-1.7B. If your workload is tool calling, structured extraction, classification, or a local assistant on a device you own, the card's warning does not apply to you and the 1 GB footprint becomes the whole argument.

Where the 12B is simply a different class of thing

Gemma 4 12B is multimodal in a way that cannot be added to LFM3-1.7B by fine-tuning. Google's card describes image and audio input projected directly into the same space as text tokens, with no separate encoders — audio natively on the 12B and its smaller siblings. LFM3-1.7B is text in, text out. Liquid's vision capability lives in LFM3-VL-2B, a separate checkpoint with a separate encoder, and a text model does not acquire one by being fine-tuned.

• Parameters — 1.73B dense for LFM3-1.7B; 11.95B dense for Gemma 4 12B. The ratio is 6.9 to 1.

• Modality — text only against text, image and audio input.

• Context — 262,144 tokens on both. This is the one row where the small model gives up nothing, and it is a real achievement of the tri-hybrid architecture rather than a coincidence.

• Footprint — about 1 GB at 4-bit against roughly 7 GB. On a phone, that difference is the difference between shipping and not shipping.

• Languages — 16 named languages against 140+ pre-trained with strong out-of-the-box support. For a Latin-script product the gap rarely bites; for a regional deployment it is decisive.

• License — the LFM Open License v1.0 is free for commercial use below $10 million in annual revenue and requires a paid license above it. Gemma 4 12B is Apache 2.0 with no revenue trigger.

The two bills are not denominated in the same currency

Gemma 4 12B has a market price because third parties serve it. Artificial Analysis reports a blended median of $0.10 per million input tokens and $0.30 per million output tokens across platforms hosting the open weights — a median across providers, not a list price, and one that moves as supply changes. At that rate, most teams never think about running a 12B themselves.

LFM3-1.7B has no such price. The Hugging Face card states plainly that the model is not deployed by any inference provider, and that is consistent with what we see on our own side: OrcaRouter does not currently route LFM3-1.7B. If you want to call it, you download 3.46 GB of BF16 weights or a roughly 1 GB 4-bit build and run vLLM, SGLang, llama.cpp, MLX or ONNX yourself. Liquid's own API will serve it, and several third-party platforms will too.

The cost comparison therefore is not tokens against tokens. It is a fine-tuning and serving bill against a per-call bill. At low volume the API wins without contest; at sustained volume on hardware you already own, a 1 GB model that leaves the rest of the GPU free is a different economic shape entirely. That asymmetry — not the four-point LiveCodeBench gap — is the real reason to consider the small model.

What each one is actually for

Choose LFM3-1.7B when the deployment target is the constraint: a phone, a laptop, a single small GPU that also has to do other work. The 256K context is genuinely unusual at this size, the 16-language vocabulary covers most regional products, and the six-way inference support means you are not betting on one runtime. Accept that you are buying a tool-calling and extraction model, not a reasoning model, and plan retrieval around the knowledge gap the vendor itself flags.

Choose Gemma 4 12B when the task needs breadth, when the input is a screenshot or a recording, or when Apache 2.0 matters to a procurement process. It scores 15 points higher on graduate science reasoning, it reads images natively, and it is available at a published median price on day one. The cost is memory: 7 GB at 4-bit is a laptop or a modest GPU, not a phone.

Choose both when you have a high-volume narrow path and a low-volume broad one. Route the broad path to a hosted generalist and keep the small model on-device for the narrow, latency-sensitive calls. OrcaRouter is built for the first half of that split — one key across more than 200 models, provider list prices passed through at 0% markup so a price change lands the same day, and automatic failover if a provider degrades. It does not help you host LFM3-1.7B, because we do not serve that model, and it would be dishonest to imply otherwise.

What would settle it

One independent evaluation of LFM3-1.7B on a harness it did not choose would do more than this entire page. Liquid's launch table is detailed and internally consistent, and this blog treats it as vendor-reported rather than verified, which is what it is. The specific things worth watching for are a third-party LiveCodeBench or SWE-bench run that reproduces the 73.46, an Artificial Analysis profile that would put both models on one scale, and any report of the model handling a real repository task — the exact use the card warns against.

Until then, the defensible reading is narrower than the headline. A 1.73B model with a 256K context, a 1 GB footprint and a strong tool-calling profile has caught a 12B generalist on one isolated coding benchmark and trails it badly on knowledge. That is not an upset. It is two models built for different jobs, and the useful question is which job you have.