What Liquid claims for LFM2.5, and which figure the card disagrees with

What was released

On October 13, 2026, Liquid AI posted LFM2.5-24B-A2B, which it calls the largest and final model in the LFM2.5 family, from a 230-million-parameter model up to this one. The post says the weights are open, base and post-trained, on Hugging Face. It says the model has 24 billion parameters in total and about 2 billion active. The model card on the same publisher's Hugging Face page says 2.45 billion active. The page filing both sentences, rather than averaging them, is Asked directly, What would replace a company benchmark? A third-party rerun of LiveClaw or AutomationBench, or a card edit that settles 2 billion versus 2.45 billion. Another writeup of the blog would not.Futureweb.

What the card describes

The card describes a 131,072-token context, 40 layers of which 30 are gated short convolutions and 10 are grouped-query attention, 64 experts with 4 active, and a vocabulary of 128,000. It lists languages that include English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish. A training-token line that reads 36 trillion in one place and about 33 trillion in another stays labeled as a card disagreement, not as a single budget. The desk isArtificial Intelligence.

What the company measured

Liquid says the model scores 48.09 against 38.19 for Qwen3.6-35B-A3B on LiveClaw, and 27.69 against 8.72 on AutomationBench, and that it is competitive on five of eight agent and tool benchmarks the company chose. It says 184 tokens per second on an Apple M5 Max, 114 on an AMD Ryzen AI Max+ 395, up to 1.7 times the next mixture-of-experts model it compared, and up to 21,000 output tokens per second on one NVIDIA B300. MarkTechPost's October 14 story relays the release. None of those speeds is an audit. What stays on the device, the standing limit under a local model, iswhat on-device inference keeps.

Why it matters on this desk

The dependency is a weight file a person can run without sending the prompt to Liquid, if the weights are what the company says they are. The day-before brief on speech, sight, and text that never leave the machine is the same dependency in a different product, Asked directly, Are the tokens-per-second figures an outside lab? No. Liquid says 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395. MarkTechPost is an account of the release, not a retest.on-device speech, sight, and OCR.

What this page will not treat as a ranking

A company table that names a rival is the company's comparison. It is not a leaderboard with a published harness and a frozen prompt set that someone else can rerun from this page. The software desk, for a version number rather than a weight file, is Asked directly, What context length and shape does the card describe? 131,072 tokens of context, 40 layers, 64 experts with 4 active, and a vocabulary of 128,000. Those are card figures, not a measured speed.Software.

What the think step is

Liquid says the chat template opens by thinking before it answers. That is a claim about the template. It does not say who can read the thinking, and it does not say the thinking is a source a reader can cite. This page will not describe the template as a control that has been tested. Asked directly, What does the company say the chat template does? It says the model thinks before it answers. That is a product claim about the template, not a proof that every answer is checked.

What to watch next

A card edit that makes the active-parameter count one number would retire the disagreement. An outside rerun of the two named benchmarks would retire the company-only label on those scores. A blog repost would not. A version string, which a weight release still needs, is Asked directly, Where do the active-parameter counts disagree? The October 13 blog says about 2 billion active. The model card says 2.45 billion active. Both are Liquid's pages. This brief does not pick one and hide the other.what a version number has to mean.

What would change the record

If Liquid withdraws the open-weight sentence, the desk changes. If the 184-token figure is marked as a preview, the speed sentence gets that label. A headline that says the model beat every rival would be wider than five of eight. The short answer this page will stand on is this: LFM2.5-24B-A2B, open weights on Hugging Face, described as 24 billion parameters with about 2 billion active. The model card also says 2.45 billion active. Liquid's speed and benchmark figures are the company's measurements, not an independent audit.

What stays the company's release

What stays specific is an October 13, 2026 Liquid release of LFM2.5-24B-A2B: open weights, 24 billion parameters, about 2 billion active on the blog and 2.45 billion active on the card, a 131,072-token context, and company benchmark and speed figures that MarkTechPost reported and did not remeasure. The disagreement on the active count is part of the record. The speeds are not an independent lab. The weights are the artifact. A later card that picks one active count is the edit this page is waiting for, and until it arrives both numbers stay on the page with the publisher that printed them.