Perceptron Mk1.5 vs LFM2.5-2.6B-Base: Rent the Perception, or Own the Checkpoint and Finish It Yourself

Both of these models are aimed at hardware you could pick up with one hand. Perceptron Mk1.5 is a perception model for embodied agents that its vendor says lands on drones, quadrupeds, smart glasses and phones. LFM2.5-2.6B-Base is a 2.69B open-weights checkpoint Liquid AI built explicitly for on-device deployment — small enough to run in under 2.5GB of memory. That shared physical ambition is the only thing they have in common, and it is a trap for anyone comparing them on a spec sheet. Perceptron Mk1.5 is a finished product you rent per token: text, images, video and audio in, and text plus machine-readable geometry out, live since September 25, 2026. LFM2.5-2.6B-Base is a raw pretrained foundation with no instruction tuning, no chat template and no benchmark claims — a checkpoint released for other people to finish. One of them answers a question today; the other is the raw material for a model that will.

The comparison only works if you name the stage

Putting a served API next to a base checkpoint is not a head-to-head, and pretending otherwise produces a nonsense table where one side "wins" every row because it has been completed by its vendor and the other has not been started. The useful question is which stage of work you are at.

Liquid AI shipped the LFM2.5 family as a set of checkpoints, and the base is the first one. Its card lists 2.69B total parameters across 30 layers — 22 double-gated short convolution blocks and 8 grouped-query attention blocks — trained on about 34 trillion tokens, with a 128,000-token vocabulary, a 131,072-token context length and support for 16 languages including English, Chinese, Japanese, Korean and Arabic. It is text-only, it is a causal language model and nothing more, and Liquid's own recommendation is blunt: this checkpoint is for tasks that require heavy fine-tuning, such as a language-specific or domain-specific assistant, training on proprietary data, or experimenting with novel post-training approaches. What turns it into the tool-calling agent that carries the family's published scores is Liquid's four-stage post-training pipeline, which the base does not include.

Perceptron Mk1.5 is at the other end of that arc. Perceptron has already done the post-training, decided the output format and priced the result — $0.15 per million input tokens and $1.50 per million output, with cached input at $0.0375. Its specifications are specific in a way the base checkpoint's are not: 36,864-token context, 8,192-token maximum output, four input modalities, a reasoning_effort control defaulting to high, function calling on chat completions, and constrained responses through JSON Schema and regex.

The reason to hold both in view is that the total cost of ownership runs in opposite directions. Mk1.5 is expensive per token and free to start. LFM2.5-2.6B-Base is free at the margin and expensive in the only currency that matters for a base checkpoint — your engineering time.

Dimension by dimension, with the roles kept straight

• What it is — Perceptron Mk1.5 is a hosted perception API with structured spatial output. LFM2.5-2.6B-Base is a downloadable 2.69B pretrained causal LM.

• Input — Mk1.5 takes text, images, video and audio (WAV, MP3, FLAC). LFM2.5-2.6B-Base takes text, and only text.

• Output — Mk1.5 returns text plus optional points, boxes, polygons, clips and timestamped <track> elements, or JSON Schema-constrained responses. The base returns token probabilities for the next token, and nothing has been tuned to make that useful.

• Standalone use — Mk1.5 works the day you get a key. LFM2.5-2.6B-Base does not answer questions in any product sense; it needs a fine-tune, and usually a chat template, before it is a model you would ship.

• Context — 36,864 tokens for Mk1.5 versus 131,072 for the base. The on-device checkpoint holds nearly four times as much text as the served perception model holds of anything.

• Footprint — Mk1.5 runs where Perceptron runs it, and its published latency work is on a single H100. The base is designed to run on a laptop, a phone or an edge box in under 2.5GB.

• Evidence — every Mk1.5 capability figure is Perceptron's own, and there is no independent index for it. The base has no benchmark claims at all: Liquid publishes scores for the post-trained LFM2.5-2.6B, not for this checkpoint.

• Licence — Mk1.5 is closed, rented per token. LFM2.5-2.6B-Base ships under the LFM Open License v1.0 with weights you download and keep.

The weights route, honestly costed

If the appeal of LFM2.5-2.6B-Base is that the weights are free, the honest version of that sentence is that the weights are the cheapest part of the project. A fine-tune on proprietary data means a dataset, a training run, an evaluation harness to tell you whether the fine-tune worked, and a serving path — Transformers, vLLM or SGLang, or one of the GGUF, ONNX and MLX builds Liquid publishes for CPU, cross-platform and Apple Silicon deployment respectively. None of that is exotic, and at 2.69B it fits on hardware that a 70B fine-tune never touches. But a base checkpoint is a starting point, and the schedule risk sits entirely on your side of the transaction.

What you get for it is control. The 16-language tokenizer and 131K context mean a domain fine-tune is not fighting the pretraining for room. The sub-2.5GB footprint means the deployment target is a device rather than a datacentre, which is the entire point of the family. And the LFM Open License lets you keep the result — a specialist no competitor can rent from the same API you can.

There is one piece of third-party evidence worth knowing before you bet on the lineage, and one gap. Liquid's post-trained LFM2.5-2.6B does have an independent measurement: Artificial Analysis lists it at an Intelligence Index of 8, ranked 9th of 49 models in its class. The base checkpoint does not — there is no page for it, which is normal for a pretrained foundation and worth stating plainly rather than treating as a verdict. What the post-trained model's index tells you is what the pipeline produces from this starting point at the small end of the index band, not what your fine-tune will produce.

The API route, and what "on-device" means for a hosted model

Perceptron Mk1.5's deployment story is more complicated than its target list suggests, and it is worth being precise. The vendor's announcement names drones, quadrupeds, smart glasses and phones as deployment targets. The model itself is served from api.perceptron.inc behind an API key, and the latency evidence Perceptron published is a median of three runs on a single H100. A single H100 is not a drone. The realistic reading is that Mk1.5 is the perception layer for a robot that has a network link or a companion compute box, not a model that runs inside the airframe — and anyone planning an offline device around it should test that assumption before designing the hardware.

What the API route buys, if the network assumption holds, is the thing a base checkpoint cannot give you at any price: coordinates. Mk1.5's <track> output carries a spatial observation and its timestamp, so a clip comes back as object positions over time rather than a description of a scene. Add the asset_idx field, which lets one request address several images or videos separately, and the request shape matches a control loop — reference frame in, geometry out, no parsing stage between the model and the code that moves something.

The costs are the ones already listed: 36,864 tokens of context, a 16,384-token per-item audio ceiling that Perceptron's docs put at roughly 21.8 minutes at about 750 tokens per minute, a reasoning_effort default of high that bills you for the most expensive reasoning path on every call you do not configure, and a set of benchmark claims produced by the company selling the model. The hand-tracking number is the one to test against your own footage: 0.9433 on egocentric hand_box against 0.4467 for Gemini 3.1 Pro in Perceptron's own run is a large margin, and large margins in vendor evaluations are exactly the ones worth reproducing.

Neither one is on our catalogue, and that changes the decision

Neither Perceptron Mk1.5 nor LFM2.5-2.6B-Base is routed on OrcaRouter today. Mk1.5 has no route at all, so reaching it means the vendor's own API — pip install "perceptron>=0.4.0" and PERCEPTRON_API_KEY — and Liquid AI has no model of any size in our catalogue, so the base checkpoint is a download from Liquid's Hugging Face rather than a call.

That matters less than it sounds, because the two are not interchangeable endpoints and no pricing argument will make them so. What a routing layer contributes here is the part of the decision that is still open. If you build a perception service on Mk1.5 and a text specialist on a fine-tuned LFM2.5-2.6B-Base, you are running two integrations and two failure modes; an OpenAI-compatible gateway in front of them turns a provider incident on either side into a fallback rather than a failed job, and a routing rule can send the vision traffic and the text traffic to different models without the application knowing which is which. One API covering 200-plus models with provider list prices passed through at 0% markup is the version of that where a vendor price change lands the same day instead of at your next renewal — though for these two specifically, today, the gateway is a planning consideration rather than an available shortcut.

What you should do, and what you should measure

If your task is "find this object in this frame and return where it is," rent Mk1.5 and test the hand-box claim on your own video before you commit a production path to it. Budget for the network assumption, set reasoning_effort deliberately instead of accepting high, and treat the 36,864-token window as the hard constraint it is. The structured output is the product, and if your downstream code can consume a coordinate without a parsing stage, that is a genuine saving in a place most vision pipelines quietly lose accuracy.

If your task is "I need a 2.6B model tuned on data nobody else has, running somewhere without a GPU," download LFM2.5-2.6B-Base and price the fine-tune honestly before you start. The checkpoint is free and the engineering is not, and the family's own lesson — that the post-trained sibling scores 8 on the independent index — is a reminder that the gap between a base and a finished model is where the actual work lives. Neither of these is the wrong answer. They are different purchases, and the only real mistake is treating a foundation you have to finish as though it were already a product.

Two models, two integrations and two failure modes is a worse position than it looks. On OrcaRouter you can set up automatic failover so a provider incident on either side degrades into a fallback instead of a failed job.