SubQ Snap 1.1 vs LFM2.5 2.6B Base: Buy the Decision or Build It from a Checkpoint

The most important word in a comparison between SubQ Snap 1.1 and LFM2.5 2.6B Base is the last one. Base means pretrained. It means Liquid AI published the checkpoint after roughly 34 trillion tokens of training and before any of the supervised fine-tuning, distillation or reinforcement learning that turns a language model into something that follows an instruction. It will not answer a question. It will not pick a label. Ask it to route a support ticket and it will continue your prompt in the style of the text it was trained on, which is not a bug — it is what a base checkpoint is for.

SubQ Snap 1.1, by contrast, is a finished product: Subquadratic's decision model, released October 22, 2026, which takes up to 384K tokens of text and images plus a fixed answer set and returns a calibrated choice. Putting them head to head as if they were alternatives is a category error, and it is a useful one, because the error is exactly the decision a team faces when it wants classification that runs on hardware it controls. The question is not "which model is better". It is "do you buy the decision, or do you build it from a starting point?"

What Liquid AI actually shipped

LFM2.5-2.6B Base arrived on August 4, 2026 with its post-trained sibling, LFM2.5-2.6B, as part of a family that runs from 230 million parameters up to a 24B mixture-of-experts. The architecture is Liquid's LFM2 design: thirty layers, twenty-two of them double-gated short-convolution blocks and eight grouped-query attention layers, a 128,000-token vocabulary, and a 128K context window added in a mid-training extension phase. At 2.69 billion parameters it fits in about 2.5 gigabytes, runs at roughly 220 tokens per second on an Apple M5 Max, around 113 on a Ryzen AI Max+, and about 30 on a phone.

The instruct sibling is where the interesting numbers live, and they are worth reading as a measure of how much post-training is worth. Liquid reports AIME25 at 51.87, LiveCodeBench v6 at 59.41, IFBench at 59.17, Multi-IF at 80.07, BFCLv4 at 56.88 and ToolSandbox at 77.83 — all vendor-reported, all on the post-trained model. The base checkpoint has no comparable instruction-following scores because it cannot follow instructions. That gap, between a checkpoint and a model you can ask a question, is four post-training stages: two rounds of supervised fine-tuning, per-domain teacher specialisation, multi-domain on-policy distillation, and agentic reinforcement learning inside real harnesses.

Liquid also says plainly what the model is not for: agentic coding, and knowledge-heavy work. That is a more honest positioning statement than most labs publish, and it is worth taking at face value.

The specs, side by side

• What it is — LFM2.5 2.6B Base is a pretrained checkpoint with no instruction tuning; SubQ Snap 1.1 is a hosted decision model with a fixed-answer API.

• Context — 128K tokens for LFM2.5 2.6B Base; 384K tokens for Snap 1.1.

• Input — text only for LFM2.5 2.6B Base; text and images for Snap 1.1.

• Output — free-form token generation, or whatever head you train on top, for LFM2.5 2.6B Base; a choice or score with a probability for Snap 1.1.

• Price — self-hosted only, no per-token bill, for LFM2.5 2.6B Base; $0.04 per 1M input tokens, output free, for Snap 1.1.

• Licence — LFM Open License v1.0, commercial use permitted with a revenue threshold, for the Liquid model; no weights at all for Snap.

• Footprint — about 2.5 GB and runnable on a phone for LFM2.5 2.6B Base; a network call for Snap 1.1.

The gap between a checkpoint and a classifier

To turn LFM2.5 2.6B Base into something that does what Snap 1.1 does, you need five things, and only one of them is the model.

You need labelled data — a few thousand examples per class is a reasonable starting point for a narrow routing task, and considerably more if the classes are imbalanced or the boundary is subtle. You need a training run: a classification head or a LoRA adapter, on top of a base checkpoint that has never seen an instruction, which means the fine-tune is doing both the task adaptation and the instruction-following work at once. You need calibration, because a softmax score is not a calibrated probability and a routing threshold set on an uncalibrated score is a threshold set by guesswork. You need an evaluation harness that you maintain, since the model will drift as your traffic changes. And you need serving infrastructure — llama.cpp, MLX, vLLM or ONNX, all of which support the architecture.

None of that is exotic. All of it is weeks of work by someone who knows what they are doing, and it recurs every time the task changes. The honest comparison is not $0.04 per million tokens against free. It is $0.04 per million tokens against four to eight weeks of engineering, plus the ongoing cost of owning a model.

The licence is not Apache

Liquid ships LFM2.5 under the LFM Open License v1.0. Commercial use is permitted, but organisations above $10 million in annual revenue need a separate commercial licence. That is a materially different proposition from Granite's Apache 2.0 or Gemma's Apache 2.0, and it is the kind of clause that gets discovered late, by legal, after the prototype works. If you are a startup below the threshold, it costs you nothing today and may cost you something later. If you are above it, price the licence before you price the GPU.

Where the base checkpoint wins outright

Air-gapped deployment is the strongest case, and it is not a close one. A 2.5-gigabyte model that runs on a phone, a laptop or a container with no network path is the only option for a field application, a classified environment, or a device that ships to customers. Snap 1.1 cannot enter any of those. It is an API, and there is no version of it that runs locally.

Volume economics are the second case, but they require real volume to matter. At two million decisions a month with 20,000 tokens of context each — 40 billion input tokens — Snap 1.1 costs $1,600 a month, or a little under $20,000 a year. A single fine-tuned 2.6B model serving that traffic on hardware you already own has no marginal cost. The break-even is somewhere around the point where the annual API bill exceeds the engineering cost of the fine-tune plus the hardware, which for most teams is a year or two out, not a month.

Full control is the third. A model you trained on your own labels does the thing you trained it to do, and nothing else. Snap 1.1 is a general decision model with a 20% accuracy figure on the Artificial Analysis Decision Index v1.0 — its index score of 65 comes mainly from abstaining on what it cannot answer, with a reported 86% non-hallucination rate. A fine-tuned 2.6B model on your domain will frequently beat that accuracy on your task, and will never abstain unless you train it to.

Where Snap 1.1 wins outright

Time to first decision is measured in minutes, not sprints. There is no training data to collect, no head to fit, no calibration curve to plot. You send a document and an option list and you get a probability back, and the probability is the product.

Long context is the second difference, and it is architectural rather than incidental. Snap handles 384K tokens in a single forward pass at constant compute per token. LFM2.5 2.6B Base has a 128K window and, as a 2.6-billion-parameter model, its effective retrieval range will be a good deal shorter than the nominal one — Liquid's own sibling model reports an AA-Omniscience score of −29.50, which is a measure of how much a small model knows, not of how well it reads, but it sets expectations.

The third is that Snap's output is already the thing you need. No parser, no schema, no retry logic for malformed answers.

Which one, then

Buy Snap 1.1 if the decision is a routing or classification call over long documents, if the data can leave your network, and if you want the capability this quarter. The build-it-yourself path is real but it is not cheaper in the first year for most teams, and the base checkpoint is the hardest possible starting point for it — you would normally begin from the post-trained LFM2.5-2.6B and fine-tune from there, not from Base.

Take the Base checkpoint if you need the model to live somewhere the network does not reach, or if you intend to own the weights permanently and have the engineering capacity to make that worth it. In that case, budget for the licence question as well as the training run.

Neither is in OrcaRouter's catalogue, and neither will be: LFM2.5 2.6B Base is a local checkpoint, and Snap 1.1 is served only through Subquadratic's API at console.subq.ai. What OrcaRouter does in a pipeline built around either is the tier that turns a decision into an action — 200+ generation models behind one key, provider list price passed through with 0% markup, and automatic failover so that a routing call landing on a model that is momentarily unavailable does not become a dropped request. The decision layer and the writing layer have different failure modes; it is worth not putting them on the same dependency.

The thing worth watching is not either model's benchmark row. It is whether a 2.6-billion-parameter checkpoint fine-tuned on a few thousand domain examples can match a purpose-built decision model on the one task a team actually has. On narrow, well-labelled classification the answer has historically been yes, and it is the reason the buy-versus-build question here is genuinely open rather than rhetorical.