LFM2.5-24B-A2B
12 workload results across 3 hardware configurations.
Fastest local config
321.3 decode tok/s
on RTX 5090 (32GB) + INTEL(R) XEON(R) GOLD 6530 (64c) + 1008GB via llama.cpp (Q4_K_M). see full run
On an RTX 5090, LFM2.5-24B-A2B (Q4_K_M) generated 308.5 tok/s on our chat-short test, and 231.9 tok/s on a 24GB RTX 4090, using 14.4 GiB and 14.2 GiB of VRAM (llama.cpp a0e6dc7, runs r_c8k6fleym09 and r_bqmpv9kncyd).
What it is
Liquid AI describes LFM2.5-24B-A2B as a mixture-of-experts model with 24B total and about 2B active parameters (Liquid's figures). We ran Liquid's own Q4_K_M GGUF (LFM2.5-24B-A2B-Q4_K_M.gguf, 14.5 GB on disk) with llama-server, all layers on the GPU.
Measured speed
| Reported hardware | Model · quant | Backend · workload | Decode (measured) | TTFT (measured) | Run |
|---|---|---|---|---|---|
| RTX 5090 (32GB) | LFM2.5-24B-A2B Q4_K_M | llama.cpp 0.6.0-dev (build 1, commit a0e6dc7) chat-short · RTX 5090 32GB, peak VRAM 14.4 GiB (8K context) | 308.5tok/s | 206ms | r_c8k6fleym09 2026-10-15 |
| RTX 4090 (24GB) | LFM2.5-24B-A2B Q4_K_M | llama.cpp 0.6.0-dev (build 1, commit a0e6dc7) chat-short · RTX 4090 24GB, peak VRAM 14.2 GiB (8K context) | 231.9tok/s | 178ms | r_bqmpv9kncyd 2026-10-15 |
| Reported hardware | Model · quant | Backend · workload | Decode (measured) | TTFT (measured) | Run |
|---|---|---|---|---|---|
| RTX A4000 (16GB) | LFM2.5-24B-A2B Q4_K_M | llama.cpp 0.7.0-dev (build 1, commit b76b133) chat-short · RTX A4000 16GB, peak VRAM 14.2 of 16.0 GiB (16K context) | 140.0tok/s | 251ms | r_j5tgvia-kjk 2026-10-15 |
Does it fit in 16GB?
Yes. On a 16GB RTX A4000 with a 16K context (Q4_K_M, llama.cpp b76b133, all layers on the GPU), it generated 140.0 tok/s on chat-short and GPU memory peaked at 14.2 GiB of the card's 16.0 GiB (run r_j5tgvia-kjk). We haven't measured other 16GB cards yet.
A user reported that a 16GB M4 Mac could not load the Q4_0 file onto the GPU (user report). We have not measured a Mac. Check your own card with the VRAM fit tool.
Versus other models on the same card
| Reported hardware | Model · quant | Backend · workload | Decode (measured) | TTFT (measured) | Run |
|---|---|---|---|---|---|
| RTX 4090 (24GB) | LFM2.5-24B-A2B Q4_K_M | llama.cpp 0.6.0-dev (build 1, commit a0e6dc7) chat-short · RTX 4090: LFM2.5-24B-A2B | 231.9tok/s | 178ms | r_bqmpv9kncyd 2026-10-15 |
| RTX 4090 (24GB) | Qwen3-Coder-30B-A3B-Instruct Q4_K_M | llama.cpp 0.6.0-dev (build 1, commit df18bee) chat-short · RTX 4090: Qwen3-Coder-30B-A3B | 232.5tok/s | 132ms | r_1k8942tw3zf 2026-10-14 |
| RTX 5090 (32GB) | LFM2.5-24B-A2B Q4_K_M | llama.cpp 0.6.0-dev (build 1, commit a0e6dc7) chat-short · RTX 5090: LFM2.5-24B-A2B | 308.5tok/s | 206ms | r_c8k6fleym09 2026-10-15 |
| RTX 5090 (32GB) | Qwen3-32B quant not reported | llama.cpp chat-short · RTX 5090: Qwen3-32B (dense) | 69.45tok/s | 327ms | r_phvxm9dcak0 2026-07-01 |
Liquid's 184 tok/s claim
Liquid quotes 184 tok/s on an M5 Max (vendor claim); we haven't measured an M5 Max, so our GPU numbers are not a like-for-like check of it. One user reported 174 tok/s on a Mac Studio (post on X, user-reported).
FAQ
How fast is LFM2.5-24B-A2B on an RTX 4090? 231.9 tok/s at Q4_K_M on chat-short (run r_bqmpv9kncyd).
How fast on an RTX 5090? 308.5 tok/s at Q4_K_M on chat-short (run r_c8k6fleym09).
How much VRAM does it need? Peak GPU memory was 14.2 GiB on the 4090 and 14.4 GiB on the 5090 at Q4_K_M with an 8K context.
Does LFM2.5-24B-A2B fit on a 16GB GPU? Yes at Q4_K_M: on an RTX A4000 16GB it ran at 140.0 tok/s on chat-short with a 16K context, peaking at 14.2 of 16.0 GiB (run r_j5tgvia-kjk). We have no 16GB Mac run.
Is it faster than Qwen3-Coder-30B-A3B? On the same RTX 4090 at chat-short they were about level: 231.9 vs 232.5 tok/s, on different llama.cpp builds.
Not measured yet
- Any Mac, including the M5 Max Liquid quotes.
- Other quants (Q5_K_M, Q6_K, Q8_0) and contexts beyond 16K.
- Answer quality. We measure speed and memory only.
Related: RTX 4090 results · RTX 5090 results · Qwen3-Coder-30B-A3B · Qwen3.6-35B-A3B · best GPU for local LLMs · why is my local LLM slow · all data
Local runs (12 workload results)
Runs from contributors' own machines via MLX, llama.cpp, vLLM, exllamav2, or ollama. Signed on the submitter's hardware.
RTX A4000 (16GB) + AMD EPYC 7402P 24-Core Processor (24c) + 126GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp@0.7.0-dev (build 1, commit b76b133) | Q4_K_M | 140.0tok/s | 453.6tok/s | 251ms | r_j5tgvia-kjk |
| chat-long | llama.cpp@0.7.0-dev (build 1, commit b76b133) | Q4_K_M | 139.9tok/s | 2,684.4tok/s | 1,223ms | r_j5tgvia-kjk |
| concurrent-decode | llama.cpp@0.7.0-dev (build 1, commit b76b133) | Q4_K_M | 142.7tok/s | no data | no data | r_j5tgvia-kjk |
| agent-trace | llama.cpp@0.7.0-dev (build 1, commit b76b133) | Q4_K_M | 139.4tok/s | 3,923.5tok/s | 391ms | r_j5tgvia-kjk |
RTX 5090 (32GB) + INTEL(R) XEON(R) GOLD 6530 (64c) + 1008GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 308.5tok/s | 553.1tok/s | 206ms | r_c8k6fleym09 |
| chat-long | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 300.4tok/s | 12,418.3tok/s | 264ms | r_c8k6fleym09 |
| concurrent-decode | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 321.3tok/s | no data | no data | r_c8k6fleym09 |
| agent-trace | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 313.2tok/s | 18,108.8tok/s | 117ms | r_c8k6fleym09 |
RTX 4090 (24GB) + AMD EPYC 7452 32-Core Processor (64c) + 1008GB
| Workload | Backend | Quant | decode tok/s | prefill tok/s | TTFT | Run |
|---|---|---|---|---|---|---|
| chat-short | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 231.9tok/s | 640.1tok/s | 178ms | r_bqmpv9kncyd |
| chat-long | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 233.4tok/s | 9,512.4tok/s | 345ms | r_bqmpv9kncyd |
| concurrent-decode | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 236.8tok/s | no data | no data | r_bqmpv9kncyd |
| agent-trace | llama.cpp@0.6.0-dev (build 1, commit a0e6dc7) | Q4_K_M | 231.0tok/s | 13,801.7tok/s | 153ms | r_bqmpv9kncyd |