LFM2.5-24B-A2B runs at 184 tok/s on an M5 Max, a published figure reported by huggingface.co (llama.cpp, Q4_K_M, 4K-token input). It needs about 17 GB of unified memory. It is Liquid AI's 24B mixture-of-experts model (2.45B active), released under the LFM Open licence.
Want to run LFM2.5-24B-A2B faster — or step up to bigger models? Find your Mac →This is a published figure: a real run, reported by its source and linked below. The index records it as given and never adjusts it.
| Metric | Value |
|---|---|
| Tokens per second | sourced184 tok/s |
| Provenance | published figure · huggingface.co |
| Quantization | Q4_K_M |
| Engine | llama.cpp |
| Context | 4K-token input |
| Figure dated | 2026-10-13 |
| Minimum RAM | 17 GB |
| Parameters | 24B MoE (2.45B active) |
| Chip bandwidth | 614 GB/s |
Each row as the index holds it, with its source. “Mac RAM” is the memory of the Mac a run used, not the model’s requirement (17 GB). The headline figure follows the reference rule: the lower-median measured plain-decode run at 5 bits or fewer (a run whose quant was not stated counts only when no stated-quant run qualifies), else the estimate; a run marked “not used to anchor estimates” is listed here but never becomes the headline.
| Speed | Quant | Engine | Mac RAM | Context | Source | Note |
|---|---|---|---|---|---|---|
| sourced184 tok/s | Q4_K_M | llama.cpp | — | 4K-token input | huggingface.co | Vendor model card: Liquid AI's Pipette device-benchmarking infrastructure on llama.cpp (build not stated). |
| community131.5 tok/s | 8-bit | MLX | 128 GB | 1k | omlx.ai | 40-core GPU. oMLX v0.7.1, MTP/DFlash off; LFM2.5-24B-A2B-MLX-8bit; 24.3 GB peak. 130.1 tok/s at 4k. Above 5 bits, so shown but never the reference. |
Ollama — install it (brew install ollama or the app from ollama.com), then pull the model by its tag, checked on ollama.com on 2026-10-24:
MLX — pick a build from Hugging Face’s MLX listings, then run it with its repo id:
Downloads: Ollama page → · model card →
The 184 tok/s figure above came from llama.cpp at Q4_K_M; other runtimes and quants run at different speeds.
Prefer a GUI? LM Studio downloads GGUF and MLX builds in one click. Step by step: install Ollama on a Mac.
| Chip | Speed | Quant | Engine | Bandwidth |
|---|---|---|---|---|
| M3 Ultra | community116 tok/s | 4-bit | MLX | 819 GB/s |
| M5 Pro | community114 tok/s | oQ4e | MLX | 307 GB/s |
| M4 Max | community90 tok/s | 5-bit | MLX | 546 GB/s |
| M4 Pro | estimated90 tok/s | Q4_K_M | Ollama | 273 GB/s |
| M6 | estimated79 tok/s | Q4_K_M | MLX | 170 GB/s |
| M5 | community74 tok/s | 4-bit | MLX | 153 GB/s |
| M4 | community59 tok/s | 4-bit | MLX | 120 GB/s |
Liquid AI's 24B MoE with 2.45B active (13 Oct 2026), a reasoning model built for on-device agents, 128K context. Liquid's card reports 184 tok/s on an M5 Max (llama.cpp Q4_K_M, 4K prompt) with 15 GB in use, so it is comfortable on a 24 GB Mac and tight on 16 GB. Community oMLX 4-bit runs measured 122 and 113.5 tok/s on an M5 Pro and 115.6 on an M3 Ultra. LFM Open License: free commercial use under $10M revenue; MLX needs mlx-lm 0.32.0+.
LFM Open licence · 24B MoE · 128K context · LLMCheck capability 35/50 · see its leaderboard rank →
LFM2.5-24B-A2B needs about 17 GB of unified memory. These current Apple Silicon Macs have the headroom to run it comfortably:
As an Amazon Associate, LLMCheck earns from qualifying purchases. These affiliate links cost you nothing extra and help keep the LLMCheck index free.
See how LFM2.5-24B-A2B stacks up against other models on your specific Mac hardware.
Open Compare Tool Full Leaderboard