otherllmRuns locally24.0B
LFM2.5-24B-A2B: Specs, License & How to Run
Open-weight 24.0B reasoning MoE (2.45B active) for on-device agents, 131K context.
- License
- other
- Parameters
- 24.0B (active 2.45B)
- HF likes
- 170 (2026-10-14)
- First seen
- 2026-10-12
TL;DR. LFM2.5-24B-A2B is a llm model from Liquid AI. Parameters: 24.0B (active 2.45B).License: other. It is listed under Closed LLM APIs.
| Developer | Liquid AI |
|---|---|
| Type | llm |
| Architecture | LFM2.5 MoE; 40 layers (30 gated short convolution + 10 GQA); 64 experts per MoE block, top-4 routing; vocab 128K |
| Parameters | 24.0B · active 2.45B |
| Modalities | text |
| Languages | English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish |
| Context | 131072 |
| License | other — LFM Open License v1.0 (license_name lfm1.0). Commercial Use rights are conditioned on You/Legal Entity not exceeding annual revenue Threshold of USD 10,000,000. Above the threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits using the work for non-commercial or research purposes are exempt from the threshold. Not legal advice. |
| Hardware | Not yet verified — Official README lists day-one GGUF, MLX, vLLM, SGLang, and ONNX checkpoints. The Liquid blog (13 Oct 2026) reports 184 tok/s on Apple M5 Max and up to 21K output tok/s on one NVIDIA B300. Those are vendor figures; forkalt has not measured speed or peak VRAM. (not independently tested) |
| Hugging Face likes | 170 (as of 2026-10-14) |
| Model card | https://huggingface.co/LiquidAI/LFM2.5-24B-A2B |
| First seen | 2026-10-12 |
LFM2.5-24B-A2B is Liquid AI’s 24B-parameter reasoning mixture-of-experts with 2.45B active parameters. Official hosts we opened on 2026-10-14: the ungated card LiquidAI/LFM2.5-24B-A2B, the base card LiquidAI/LFM2.5-24B-A2B-Base, the LFM Open License v1.0, the launch post LFM2.5-24B-A2B: Built for Fast On-Device Agents (13 October 2026), and docs.liquid.ai.
This identification page hangs off open alternatives to closed LLM APIs. That mapping is editorial for local / on-device agent checkpoints. We have no sourced quality comparison against Claude, GPT, or any other named closed API.
Do not confuse this checkpoint with the later LFM3 family (~2B text/VL/audio/omni; announced 27 October 2026) — LFM3 vs LFM2.5-24B-A2B. Do not confuse it with LiquidAI/d1-3B, which remains on the classification watch list.
What we verified (2026-10-14)
| Field | Value | Source |
|---|---|---|
| Hugging Face likes | 170 | HF API LiquidAI/LFM2.5-24B-A2B |
| HF downloads (monthly counter) | 2,565 | HF API |
| Card created | 2026-10-12T05:35:46Z | HF createdAt |
| Card last modified | 2026-10-13T16:59:33Z | HF lastModified |
| Gated | false | HF API |
| License | other / lfm1.0 (LFM Open License v1.0) |
HF cardData + LICENSE |
| Total / active params | 24.0B / 2.45B | HF README “Model Details” |
| Layers | 40 (30 gated short conv + 10 GQA) | HF README |
| Experts | 64 per MoE block, top-4 | HF README + blog |
| Vocab / context | 128,000 / 131,072 | HF README |
| Languages | EN, AR, ZH, FR, DE, IT, JA, KO, PT, ES, VI, TH, ID, HI, RU, PL | HF README |
| Blog shorthand | “2B active” | Liquid blog — we prefer the card’s 2.45B |
How to run (from the official README and blog)
The README lists native, FP8, GGUF (llama.cpp), ONNX, and MLX-4bit siblings. The blog names day-one llama.cpp, MLX, vLLM, SGLang, and ONNX. Docs: lfm25-24b-a2b. forkalt has not published a tested serve recipe.
Vendor speed claims (self-reported, blog): 184 tok/s on Apple M5 Max; up to 21K output tok/s on one NVIDIA B300. Not our stopwatch.
Vendor agent benches (self-reported)
The Liquid blog Table 1 (13 Oct 2026) reports LiveClaw 48.09 versus Qwen3.6-35B-A3B 38.19, and AutomationBench 27.69 versus 8.72. Those rows are self-reported. The same table includes other open MoEs. It is not a Claude or GPT comparison. The card also embeds an official benchmark figure.
License
LICENSE is LFM Open License v1.0. Licensor is named Liquid AI, Inc. Section 5: Commercial Use rights are conditioned on You or Your Legal Entity not exceeding the Threshold of USD 10,000,000 annual revenue. Commercial Use by an entity over that threshold is not licensed under this agreement. Qualified Non-Profit Organizations using the work for Non-Commercial or Research Purposes are exempt from the threshold. Not legal advice.
Related chat on this site: LFM3 (Liquid AI ~2B family; not this 24B/2.45B MoE) — LFM3 vs LFM2.5-24B-A2B — Spark-X2.6-4B (XHToken Apache-2.0 ~4B on-device agent; not Meta Muse Spark) — Spark-X2.6-4B vs LFM2.5-24B-A2B — DeepSeek-V4.1-Flash on ChatGPT alternatives, and Kolibri-1 on the same closed LLM APIs hub. Runtimes: vLLM, DeepInfer vs vLLM. Catalog: home. Editorial: About. Pixel editors: PhotoCraft vs GIMP.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| LiveClaw (vendor table) | 48.09 (self-reported) | https://www.liquid.ai/blog/lfm2-5-24b-a2b |
| AutomationBench (vendor table) | 27.69 (self-reported) | https://www.liquid.ai/blog/lfm2-5-24b-a2b |
| Decode speed on Apple M5 Max (vendor) | 184 tok/s (self-reported) | https://www.liquid.ai/blog/lfm2-5-24b-a2b |
Deploy options
Comparisons
FAQ
Are the weights public?
Yes. LiquidAI/LFM2.5-24B-A2B is on Hugging Face (created 2026-10-12T05:35:46Z; not gated; 170 likes and 2,565 downloads in the 2026-10-14 API check). A base checkpoint is LiquidAI/LFM2.5-24B-A2B-Base. The Liquid blog is dated 13 October 2026.
How big is LFM2.5-24B-A2B?
The Hugging Face README (checked 2026-10-14) lists 24.0B total parameters and 2.45B active, 40 layers (30 double-gated short convolution blocks + 10 GQA), 64 experts with top-4 routing, a 128,000-token vocabulary, and 131,072 context. The blog shorthand “2B active” is coarser than the card; we use 2.45B.
Can I use it commercially?
The LICENSE file is LFM Open License v1.0. Section 5 conditions Commercial Use on You/Legal Entity not exceeding USD 10,000,000 annual revenue. Above that threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits for non-commercial/research purposes are exempt from the threshold. Not legal advice.
Does it beat Claude or GPT?
We do not claim that. This page identifies an open-weight local agent LLM. Blog tables versus Qwen3.6-35B-A3B and other open MoEs are vendor self-reported. There is no sourced bake-off against named closed APIs on this page.
Sources & verification
- https://huggingface.co/LiquidAI/LFM2.5-24B-A2B
- https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-Base
- https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/raw/main/LICENSE
- https://www.liquid.ai/blog/lfm2-5-24b-a2b
- https://docs.liquid.ai/lfm/models/lfm25-24b-a2b
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.