HiRO-ACEv2 vs LFM2.5 2.6B Base: Neither One Is a Product, and They Are Not Unfinished in the Same Way

Put HiRO-ACEv2 and LFM2.5 2.6B Base side by side and the first thing you notice is that neither is something a developer can call. HiRO-ACEv2 is Ai2's second-generation climate emulator, uploaded as a 10.9 GB Hugging Face repository on 12 October 2026 with no announcement and a preprint dated the following day; LFM2.5 2.6B Base is Liquid AI's 2.69-billion-parameter pretrained checkpoint, created on 1 August 2026, explicitly described by its own vendor as the checkpoint you reach for only when you need substantial adaptation. Both are raw material. Both are downloadable today. And the ways in which each one is not a finished product are close to opposites, which is the only reason this comparison is worth writing rather than dismissing.

The Base checkpoint is unfinished because it has not been taught to follow instructions yet. HiRO-ACEv2 is unfinished in the sense that it never will be a general-purpose model — it is an application that arrived already trained, already specialised, and already wrapped in stated limits. One is a substrate waiting for a purpose. The other is a purpose with no substrate for anything else.

What each repository actually contains

Liquid AI's Base repository is a model. The card lists 2.69 billion parameters across 30 layers — 22 double-gated short convolution blocks and 8 grouped-query-attention layers, a hybrid arrangement the LFM2 family is built around — with a 131,072-token context window, a 128,000-entry vocabulary, 34 trillion tokens of pretraining, and coverage of 16 languages including English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish. It is text-only, and it is not instruction-tuned. Download formats run from the native Transformers checkpoint through GGUF for llama.cpp, ONNX for cross-platform deployment, and MLX for Apple silicon; the Q4_K_M quantisation lands around 1.67 GB.

Ai2's repository is not a model in that sense. It holds two checkpoints — HiROv2.ckpt at about 1.28 GiB and ACE2S-X-SHiELD.ckpt at about 2.98 GiB — plus ten years of NetCDF forcing data at roughly 609 MB per file, two initial-condition files, three YAML configs and a shell script. The roughly 4.3 GiB of actual weights is a minority of the 10.9 GB total; most of what you download is the input data window the framework was built to consume. There is no quantisation, no alternative runtime, no ONNX export, and no smaller variant. There is one way to run it, and the README describes it: install Ai2's fme package, edit four paths in the config, execute the script.

Two different kinds of licence problem

The licence tags are where the contrast becomes practical rather than philosophical, and where this pair genuinely differs from a model like the Apache-2.0 releases you might expect to find here.

• HiRO-ACEv2 — Apache 2.0, with Ai2's Responsible Use Guidelines attaching research and educational intent to the grant, and training data attributed to X-SHIELD, provided by GFDL.

• LFM2.5 2.6B Base — the LFM Open License v1.0, not an OSI-style permissive licence. Its terms live in the repository's own LICENSE file, and the commercial conditions are the reason to read that file rather than the tag before you build a product on the checkpoint.

• What that means for the fork you are about to make — with HiRO-ACEv2 the question is whether your use is research; with LFM2.5 2.6B Base the question is whether your use is commercial, and at what scale. Different questions, different answers, and neither is answered by the word "open."

The output shapes have nothing in common, and that is the useful part

The Base checkpoint produces text. Token by token, autoregressively, with the usual sampling surface — the card suggests temperature 0.6 and top-p 0.95 for the post-trained sibling's generation style, and a base checkpoint asks you to decide that yourself. Everything downstream of it is a language task: continue a document, classify a span, translate, emit structured data, or serve as the starting point for a fine-tune that teaches it a domain or a language it does not currently cover well.

HiRO-ACEv2 produces gridded physical fields at 3 km resolution — six-hourly averaged surface precipitation rates, instantaneous 10-metre eastward and northward winds, and sea level pressure — downscaled from 100 km input by a diffusion model, with the coarse state itself generated autoregressively by a separate emulator of X-SHIELD. Nothing about that pipeline is text. The evaluation that matters is whether the distribution of the output matches a reference simulation out to its far tail, and the paper behind the release is about exactly that: the authors report tracking reference precipitation and near-surface wind distributions to the 99.9999th percentile, and broadly following the reference tropical cyclone intensity distribution, with a claimed inference cost more than 300 times below a conventional numerical global storm-resolving model.

Both of those are vendor or author claims, unreproduced. The difference is what a reproducer would have to do. Verifying a claim about the Base checkpoint means a task suite and a checkpoint's worth of GPU hours. Verifying a claim about HiRO-ACEv2 means a ten-year forcing dataset, a reference simulation to compare against, and a distributional analysis in the far tail — which is why the second claim is likely to sit unchallenged for considerably longer.

What you would actually build on each

The Base checkpoint's natural use is adaptation, and Liquid AI is unusually direct about it: language- or domain-specific assistants, training on proprietary data, or testing new post-training methods. That is a checkpoint whose value arrives after your work, not before it. The post-trained sibling — LFM2.5 2.6B, created three days earlier on 28 July — is the one built for agentic workloads, and the download figures say everything about which of the two the community wanted: the post-trained model had passed 114,000 downloads and 823 likes by mid-October, against 6,282 downloads and 51 likes for the Base. The Base is a specialist's tool and the market has priced it accordingly.

HiRO-ACEv2's natural use is running an ensemble. The framework's selling point is not one forecast but many: the paper's motivation is generating large ensembles cheaply enough to study tropical cyclone behaviour and extremes at 3 km, over multi-decade periods and arbitrary global regions. You do not fine-tune it, and the model card does not invite you to. You run it, and you either trust its limits — nine years of training data, no applicability outside 66°S to 70°N, patch-boundary artifacts above the 16-by-16 one-degree training tile — or you do not.

Put differently: one of these models is a starting point you modify, and the other is an instrument you operate. Reading them as interchangeable "open models" because both are downloadable is the mistake this page exists to head off.

The one thing they do share: neither is billable per token

Where the two genuinely converge is cost shape. Neither has a rate card. LFM2.5 2.6B Base is free to download and cheap to run precisely because it is small — a couple of gigabytes in bf16 and well under two gigabytes quantised, which is the entire point of the LFM2.5 design and the reason it exists alongside a post-trained sibling tuned for on-device work. HiRO-ACEv2 is cheap for a different reason: it is a narrow emulator, and its economics come from replacing an enormously expensive numerical simulation rather than from fitting on a phone.

So if your workload is text, the Base checkpoint is a question about fine-tuning, and the everyday version of that workload — the one where you want a model to answer rather than a checkpoint to train — looks like a call, not a download. That is the layer OrcaRouter sits in: one API key across two hundred-plus hosted models, with provider list price passed through at no markup and automatic failover between providers when one degrades. Neither model in this comparison is in that catalogue, and neither could be in its present form; LFM2.5 2.6B Base is a checkpoint rather than a service, and HiRO-ACEv2 has no request-shaped interface at all. What we route is everything around them — the summarisation, the extraction, the harness code and the agent loop that turns a downloaded artifact into a pipeline.

Picking, in practice

If you have data in a domain or a language the base model has not seen enough of, and you have the compute to post-train, LFM2.5 2.6B Base is a cheaper starting point than most of its competition — read the licence file before you read the benchmark you were hoping for, because on this checkpoint the licence is the document that decides whether your project is allowed. If you want a small agentic model to run on a device, you want the post-trained sibling rather than the Base, and the download numbers suggest the community reached that conclusion before you did. If you want a 3 km atmospheric field over three decades of simulated climate, neither of these is relevant and only one file in this article is.

The useful thing to take away is that "open weights" has stopped being a useful category on its own. A base checkpoint, a post-trained sibling and a research emulator can all carry a permissive-sounding tag and mean three entirely different things by it.

Comparisons like this one are also where the failure modes show up: 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.