LFM2.5-24B-A2B on iPhone: Liquid AI's new on-device agent model, and why no iPhone can hold it
Liquid AI shipped LFM2.5-24B-A2B on October 13 as a 24B mixture-of-experts model for on-device agents — but "on-device" means a Mac or a mini PC here, not a phone. Even Q4_K_M is 14.53 GB, nearly three times the 5.07 GB ceiling of the biggest iPhone.
On October 13, Liquid AI released LFM2.5-24B-A2B, the largest model in its LFM2.5 family and the one it is pitching hardest at "on-device agents" — tool use, instruction following, and long-running agentic tasks running locally instead of through an API. Both a base checkpoint and a post-trained one are open-weight, with day-one support across llama.cpp, MLX, vLLM, SGLang and ONNX. That is a real, notable release from a vendor this blog already tracks.
Short answer for this one: no iPhone can run it, and it isn't close. Even the smallest file in Liquid's own GGUF release, Q4_0, is 13.58 GB — about 2.7 times the 5.07 GB ceiling of the 12 GB iPhone 17 Pro, the most memory any current iPhone has. "On-device" in Liquid's own materials means a laptop or a mini PC, not a phone, and the model card's own numbers make that clear once you look at where it was benchmarked.
What actually shipped
LFM2.5-24B-A2B is a mixture-of-experts model: 24.0B parameters total, with about 2.45B active per token through 64 experts and top-4 routing. It keeps the hybrid LFM2 architecture — 40 layers split into 30 gated short-convolution blocks and 10 grouped-query-attention blocks — and extends two things from the earlier LFM2 family: context grows from 32,768 to 131,072 tokens, and the vocabulary doubles from 65,000 to 128,000 tokens. Pre-training reportedly scaled to around 33 trillion tokens, with a reasoning-weighted mid-training stage at a 32K context followed by a shorter stage at the full 131K context. License is Liquid's own LFM Open License v1.0, the same custom (not Apache/MIT) terms it has used for earlier LFM2 releases, which cap commercial use free of charge below a revenue threshold and require a separate agreement above it.
Why "2.45B active" doesn't change the phone math
The pitch for mixture-of-experts models is speed: only a fraction of the parameters run on any given token, so inference is fast relative to the total parameter count. That is true for compute. It is not true for memory. Every expert can be selected depending on what the model is routing on that token, so the device still has to hold all 24B parameters resident to serve the model at all — the active-parameter number describes the compute path, not the RAM footprint. This is the same lesson our RAM post makes about dense models: the ceiling is set by file size, not by how much of the model any one answer actually touches.
Does it fit your phone
Liquid's own GGUF repository (LiquidAI/LFM2.5-24B-A2B-GGUF on Hugging Face) lists the file sizes: Q4_0 is 13.58 GB, Q4_K_M is 14.53 GB (about 4.85 bits per weight) and Q8_0 is 25.49 GB.
Running those against the same ceilings used throughout this blog (55% of physical RAM, with a 20% safety margin and 600 MB reserved for everything besides the weights — full math in the RAM post):
| Device | Weight ceiling | Q4_0 (13.58 GB) | Q4_K_M (14.53 GB) | Q8_0 (25.49 GB) |
|---|---|---|---|---|
| 4 GB — iPhone 12/13 | 1.29 GB | ✗ | ✗ | ✗ |
| 6 GB — iPhone 12–14 Pro, 14, 15 | 2.23 GB | ✗ | ✗ | ✗ |
| 8 GB — iPhone 15 Pro, 16, 16e, 17 | 3.18 GB | ✗ | ✗ | ✗ |
| 12 GB — iPhone 17 Pro, 17 Pro Max, Air | 5.07 GB | ✗ | ✗ | ✗ |
Unlike the Bonsai 2 27B post, where a ternary 27B model missed the top iPhone by under a gigabyte, this isn't a near-miss. Even for Q4_0, the gap between 13.58 GB and 5.07 GB is bigger than the entire weight budget of a 12 GB iPhone.
The benchmark claims
Liquid's own launch comparison runs LFM2.5-24B-A2B against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning across sixteen benchmarks. By the secondary coverage we could read, Liquid's model leads on agentic and instruction-following tasks: AutomationBench 27.69 vs. Qwen's 8.72, LiveClaw 48.09 vs. 38.19, ToolSandbox 80.71 vs. 75.02, IFEval 96.75, IFBench 74.69, Multi-IF 84.38 (all ahead of the comparison models), and a near-tie on BFCLv4 (65.01 vs. 65.00). Qwen3.6-35B-A3B leads on coding and knowledge: SWE-Bench Verified 66.20 vs. 51.00, GPQA Diamond 83.69 vs. 74.42, PinchBench 89.93 vs. 83.18, and WebArena Verified 37.62 vs. 26.75. These are Liquid's own numbers on its own test setup; one independent write-up specifically flagged that the Gemma and Nemotron columns look like figures Liquid reproduced rather than numbers those vendors measured themselves. No independent benchmark run of the new checkpoint had surfaced at the time of writing — treat all of it as a vendor's comparison table, not a settled result.
Liquid's own speed numbers are also laptop- and workstation-class: about 184 tokens/second on an Apple M5 Max and 114 tokens/second on an AMD Ryzen AI Max+ 395 mini PC, with up to 21,000 output tokens/second on a single Nvidia B300 GPU at high concurrency. None of that hardware is a phone, and nothing in Liquid's own materials claims it should be.
Sources
Checked on 14 November 2026 against the primary sources: the model's config.json on Hugging Face (40 layers, 64 experts with top-4 routing, 131,072-token context, 128,000-token vocabulary, lfm1.0 license), its parameter count (23.97B), Liquid's GGUF file list for the sizes above, and Liquid's own announcement, dated October 13. The benchmark scores and speed figures are Liquid's, taken from its launch comparison as reported in launch coverage; they were not re-measured here.
Should you download it
If you have an Apple Silicon Mac, a capable Windows mini PC, or a server GPU, LFM2.5-24B-A2B is a legitimate, fast local agent model and the numbers above are worth your attention. If you only have an iPhone, this one isn't for you — not this version, and not a smaller quantization of it either, since even Q4_K_M is nearly three times over budget on the biggest iPhone made. If you want Liquid's models on a phone, the smaller members of the same family do fit: Pocket AI offers LFM2.5 230M (146 MB), 350M (219 MB), 1.2B (697 MB), 2.6B (1.6 GB) and 8B-A1B (4.8 GB, 12 GB iPhones only). For everything else that runs today, see which AI models run on your iPhone.