ZwZ2-9B vs LFM2.5 2.6B Base: A Category Error Worth Understanding

Putting ZwZ2-9B next to LFM2.5 2.6B Base looks like a size comparison, and it is not one. ZwZ2-9B is a 9.41-billion-parameter vision-language checkpoint that inclusionAI posted on 16 October 2026 with no model card; LFM2.5 2.6B Base is a 2.697-billion-parameter text-only pre-trained model that Liquid AI published on 1 August 2026, with a card that says in its own words it is the checkpoint "used to create all the LFM2.5-2.6B variants". One is a thing you adopt. The other is a thing you build on. Choosing between them is not like choosing between two cars — it is like choosing between a car and an engine block, and the only useful question is which of those you actually needed.

So this piece does two things. It gives you the real numbers on both, because you came here looking for a comparison and the numbers are worth having. And then it explains the distinction that the comparison hides, because getting that wrong is how teams end up three weeks into a fine-tune they did not need, or running a 9.4B server model for a job a 3B laptop model would have handled.

What each of these artefacts actually is

LFM2.5 2.6B Base is deliberately unfinished. It is the pre-trained checkpoint at the root of Liquid AI's LFM2.5-2.6B family — 30 layers alternating between double-gated short-convolution blocks and grouped-query attention, 34 trillion tokens of pre-training, a 128,000-token vocabulary and a 128,000-position ceiling. It is text-only. It is designed for on-device deployment, which is the phrase Liquid AI uses in the card, and the whole point of the "Base" suffix is that the post-trained conversational variant and the agentic variant are separate downloads derived from it. You are meant to fine-tune it or distil into it. Out of the box it is not an assistant.

ZwZ2-9B is the opposite kind of object. It is a finished-shaped checkpoint that has not been described. Its config carries a 27-layer vision tower, image and video token ids, a video processor capped at 768 frames, a chat template with a baked-in default system prompt, and 24 of its 32 layers running linear attention instead of a growing key-value cache. That is a deployment target, not a training substrate. What is missing is everything a purchaser would ask for: no licence field, no evaluation results, no statement of what it was trained on, and no entry in any of inclusionAI's ten public collections.

• Parameters — ZwZ2-9B: 9,409,813,744. LFM2.5 2.6B Base: 2.697 billion.

• Purpose — a deployment-shaped multimodal checkpoint vs a pre-trained base for fine-tuning.

• Modalities in — text, image and video vs text only.

• Context ceiling — 262,144 positions vs 128,000 (Liquid AI's card rounds this to 131,072).

• Vocabulary — 248,320 vs 128,000.

• Attention design — 24 linear layers plus 8 full-attention layers vs 22 convolution layers plus 8 attention layers.

• Licence — none declared vs the LFM Open License v1.0, which is a custom licence rather than Apache-2.0.

• Documentation — none vs a model card, a technical report and a stated training budget.

Two lines there are more interesting than they look. Both models land on exactly 8 attention layers, and both fill the rest with something cheaper — ZwZ2-9B with recurrent linear attention, LFM2.5 2.6B Base with short convolutions. That convergence is not a coincidence; it is what "efficient architecture" has meant for the last two years, arrived at from opposite directions and at opposite scales. And the context ceilings are not a contest: 262,144 against 128,000 says almost nothing, because the 2.7B model was built to sit on a laptop where a quarter-million-token cache is not a realistic working set.

The word "Base" is the whole decision

If you take one thing from this pairing, take this. A base checkpoint has no instruction following, no chat formatting worth relying on, and no safety tuning — and that is a feature if you are building a domain model, because it means you are not fighting someone else's alignment when you train yours. It is a defect if you wanted an assistant, because you will spend your first two weeks rediscovering that pre-trained models complete text rather than answer questions.

ZwZ2-9B sits on the other side of that line. Its chat template injects a default system prompt when none is supplied, and it emits an empty <think></think> block unless you explicitly request reasoning — the opposite default from its Qwen3.5-9B parent. Somebody post-trained it for one-shot answering and then published it without saying so. That is a real signal, and it is the only signal in the repository that describes intent.

The deployment maths, which is where the real difference lives

A 2.697B text model and a 9.41B vision model are not alternatives on the same hardware. The LFM2.5 base is built for phones, laptops and single-board machines; Liquid AI's framing is on-device inference, and a 2.7B checkpoint quantised to four bits is a few gigabytes. ZwZ2-9B keeps a 27-layer vision tower alongside 32 language layers and roughly 2 GB of embedding matrix in each direction — the vocabulary alone is a 248,320-by-4,096 table — so it belongs on a server, or at least on a desktop with a real accelerator.

That is why the honest architecture for most teams is not "choose one". It is two tiers. Put the small documented model where latency and privacy matter and the task is narrow, and route the hard cases to a hosted model where the variable cost only appears when the hard case does. Getting that split right is more valuable than any single model choice on this page, and it is also the thing that is hardest to change later, because it is the one that touches your code.

Doing it behind a single interface is the point of a router. OrcaRouter exposes 200-plus models from one OpenAI-compatible endpoint and one API key, with automatic failover, so the cheap tier and the expensive tier are a model string apart rather than two integrations — and provider list prices pass through at 0% markup, so running the expensive tier is the same cost as it would be anywhere else. We do not host LFM2.5 2.6B Base and we do not host ZwZ2-9B, so neither of these is a call we can make for you — Liquid AI distributes its own weights through the vendor's own channels, and inclusionAI has published no endpoint at all. What we can put on the hosted side of that split today is qwen/qwen3-vl-8b-instruct, an 8B vision-language model with a 131K-token context at $0.18 per million input tokens and $0.70 per million output, which is the closest documented stand-in for the small multimodal tier.

Licensing is where these two diverge most sharply

Neither licence is Apache-2.0, and the reasons are different in a way that matters. LFM2.5 2.6B Base ships under the LFM Open License v1.0, which is a custom licence — it is published, it is readable, and it is the same licence the rest of the family uses, which means a legal review you do once covers every LFM2.5 model you touch. ZwZ2-9B declares nothing. A missing licence is not a permissive licence; absent an explicit grant, the conservative reading is that no rights have been extended, and that is a materially worse position than a custom licence you can read and evaluate.

For a base checkpoint you intend to fine-tune and possibly redistribute, this is not a footnote. Derivative-work terms are exactly the clause that custom licences tend to have opinions about, and the absence of any clause at all does not resolve in your favour.

How to choose, stated plainly

Choose LFM2.5 2.6B Base if you have a narrow text task, a target device, and the appetite to run your own training — and accept that you are the one supplying the instruction-following. Its documented training budget and stated licence mean you know what you are starting from.

Choose ZwZ2-9B if you need images or video in the input, you have the GPUs to serve a 9.4B model, and you are willing to be the first organisation on earth to evaluate it. That last clause is not rhetorical. Before committing, the questions to answer are whether it accepts your image formats, whether it returns the output shape you need with thinking disabled, and which licence you would actually be operating under. Two of those three have answers in the repository. The third does not.

Benchmarks age and prices move; the failure record does not: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.