LFM3-1.7B-Base vs Gemma 4 12B: A Specialist Trunk Against a Generalist

Put the two downloads side by side and the first difference is not a benchmark. LFM3-1.7B-Base, published to Hugging Face on October 26, 2026 and named in Liquid AI's family launch the next day, is a 3.46 GB bfloat16 file containing a 1.73-billion-parameter pre-trained text model. Gemma 4 12B, released by Google DeepMind on June 3, 2026, is an 11.95-billion-parameter dense model that ships with instruction tuning, a 256K context, and the ability to read images and audio as well as text. One is a component. The other is a product. Comparing them on capability is a category error, and comparing them on price per token is nearly as bad, because the base checkpoint has no token price — it has a fine-tuning bill. Routing does not add a margin to a provider's list price. 0% markup See the pass-through pricing on our pricing page.

The comparison earns its page anyway, because both are open-weight models a small team might download this week for the same underlying reason: they want a model they control, on hardware they own, without a per-call contract. The useful question is which kind of asset gets you there — a small trunk you shape, or a large generalist you adopt.

The benchmark comparison that cannot be made

LFM3-1.7B-Base has no published evaluation. Not a thin one, not a vendor table with an asterisk — none. The model card gives specifications and a fine-tuning guide, and Liquid's launch post carries a substantial benchmark table for LFM3-1.7B, the post-trained sibling that was built from this checkpoint. Those are different artifacts. A base model predicts text; a post-trained model follows instructions. Scoring the first on an instruction benchmark tells you about the second.

Gemma 4 12B, by contrast, arrives with a full family table for the instruction-tuned checkpoints: 77.2% on MMLU Pro, 78.8% on GPQA Diamond, 77.5% on AIME 2026 without tools, 72.0% on LiveCodeBench v6, a Codeforces Elo of 1659, 69.1% on MMMU Pro, and 43.4% on MRCR v2 at 128K with eight needles. All of those are Google's own figures and none has been independently reproduced. Artificial Analysis separately places Gemma 4 12B at an Intelligence Index of 14, ranked 20th of 141 models in its comparison class, with a measured output speed of 110.9 tokens per second.

So the honest scoreboard has one populated column and one empty one, and the empty column is not a failure of the model. It is what a base checkpoint is. Anyone who puts 49.37 — Liquid's own average for the tuned LFM3-1.7B — next to Gemma's MMLU Pro is comparing two different suites, two different models and two different levels of tuning.

What each checkpoint is actually for

LFM3-1.7B-Base exists to be changed. Liquid's card says it is recommended only for heavy fine-tuning — language-specific assistants, domain-specific assistants, training on proprietary data, novel post-training approaches — and points anyone who wants chat or tool use at the tuned model instead. It was pre-trained on 22 trillion tokens, the last 3.5 trillion of them curated mid-training data with agentic traces and tool calls, and that stage extended the context to 262,144 tokens. At 4-bit it takes about 1 GB, which is the entire design argument: a trunk small enough to fine-tune on one GPU and run on a phone.

Gemma 4 12B exists to be used. It is the mid-size dense member of Google's five-model Gemma 4 family, with no separate multimodal encoders — image patches and audio waveforms are projected directly into the same space as text tokens. It runs on a laptop in roughly 7 GB at a 4-bit quantization, carries Apache 2.0 with no revenue trigger, and has the kind of ecosystem support that a model six months old accumulates: quantized builds, framework integrations, and a community that has already found its failure modes.

Six differences that decide it

• Parameters — 1.73B dense against 11.95B dense. The size ratio is 6.9 to 1, and it drives everything downstream: memory, fine-tuning cost, inference speed and the ceiling on general knowledge.

• Modality — text in, text out, for LFM3-1.7B-Base. Text, image and audio in, text out, for Gemma 4 12B. The base checkpoint cannot see a screenshot, and no amount of fine-tuning adds a vision encoder that was never trained.

• Context — both are 262,144 tokens. This is the one row where the small model gives nothing away, and it is a genuine achievement of the hybrid architecture rather than a marketing coincidence.

• Languages — sixteen listed for LFM3-1.7B-Base, including Arabic, Hindi, Thai, Vietnamese and Indonesian. Gemma 4 12B is pre-trained on 140+ languages with out-of-the-box support claimed for 35+, a far wider net and a real advantage for multilingual deployment.

• Vocabulary — 128,000 tokens for LFM3-1.7B-Base, 262,000 for Gemma 4 12B. The larger vocabulary compresses text more aggressively in Latin-script languages and in code; the smaller one keeps the embedding matrix small, which matters when the whole model is 3.46 GB.

• License — the LFM Open License v1.0 is free for commercial use below $10 million in annual revenue, with a commercial license above that threshold. Gemma 4 12B is Apache 2.0, with Google's acceptable-use terms alongside the weights rather than as a revenue condition.

The one same-suite number, and why it does not transfer

There is exactly one place where Liquid has published a score against Gemma 4 12B on the same evaluation suite, and it involves a different model. Across 62 vision benchmarks, Liquid reports LFM3-VL-2B, its 1.9B vision model, at 66.0 on the suite's general category against 74.0 for Gemma 4 12B. Elsewhere in the same table the smaller model is closer, and Liquid's summary is that the vision model "matches" Gemma 4 12B despite being more than six times smaller.

Read the table rather than the sentence and it says something more precise: a 1.9B multimodal model gets within eight points of a 12B multimodal model on a broad vision suite, and the gap is uneven across categories. Read it as a statement about LFM3-1.7B-Base and it says nothing at all. The text base checkpoint has no vision tower and was not evaluated on any of those benchmarks. The claim belongs to a sibling, and it is the kind of provenance slip that turns a careful comparison into a misleading one.

What the 12B generalist is worth in money

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 that host the open weights — a median across providers, not a list price, and one that will move as the supply side changes. That is cheap enough that most teams never think about running it themselves.

LFM3-1.7B-Base has no such price because it is not a served model. Its cost is the fine-tuning run, the data pipeline and the serving stack, and those numbers vary by more than an order of magnitude depending on what you are building. Liquid's pitch is that the run fits on a single GPU and the result fits in about a gigabyte, which is true of the model and says nothing about the cost of the labelled data. A specialist fine-tune is not cheaper than calling a generalist API unless you already have the data; if you do, it is dramatically cheaper at volume.

This is where a routing layer earns its place. OrcaRouter passes provider list prices through at 0% markup, so a vendor price cut lands the same day rather than waiting for a repricing cycle, and the same key reaches more than 200 models with automatic failover. The practical pattern for a team weighing this decision is to keep the generalist on the routed path while the specialist fine-tune is being evaluated, then shift traffic when the fine-tune wins on your own held-out set — not on a vendor average.

Three buyers, three answers

You have proprietary data and a narrow task. Take LFM3-1.7B-Base. The 1.73B trunk is small enough to fine-tune on one GPU, the 256K context survives the fine-tune, and the sixteen-language vocabulary covers most regional deployments. Gemma 4 12B can also be fine-tuned, but you would be paying to retrain capability you are about to specialize away.

You need a model that works this afternoon. Take Gemma 4 12B. The base checkpoint cannot answer a question, and the tuned sibling is a different download with a different license question. Gemma 4 12B is instruction-tuned, multimodal, documented, and available at a published median price. The time-to-first-answer is measured in minutes.

You need both. This is the common case and the one the framing hides. Fine-tune the trunk for the narrow, high-volume task and route everything else to a generalist. The two are not substitutes at the architecture level, and treating the decision as exclusive is how teams end up either over-paying for a specialist or under-serving a use case their small model was never trained for.

The license is the sleeper issue

Apache 2.0 is the reason many companies stop reading at Gemma's license. There is no revenue threshold, no commercial trigger and no negotiation — the weights are yours to modify and serve, subject to Google's acceptable-use policy. Liquid's LFM Open License v1.0 is also permissive at small scale, but it defines a $10 million annual revenue threshold above which a commercial license applies. For a startup the difference is theoretical. For a business unit inside a larger company, it is a procurement conversation that starts before the first fine-tune does.

What would change the answer

Two pieces of evidence would move this comparison from a category argument to a numerical one. The first is a base-model evaluation from Liquid — even a small one, on a held-out corpus — that lets a reader judge whether 22 trillion tokens and the tri-hybrid stack produced a better starting point than LFM2.5-2.6B-Base. The second is an independent fine-tune of LFM3-1.7B-Base on a public task, reported against a Gemma 4 12B baseline. Neither exists yet, and the launch post does not promise them.

Until they do, the decision rests on what you are buying rather than which model is better: a 3.46 GB trunk with no scores and no instruction tuning, or a 12B generalist with a full table of vendor numbers and a license you never have to revisit. That is not a close call on capability. It is a close call on fit, and the empty column in the scoreboard is the most honest thing on this page.