All models › LiquidAI

LFM2.5-24B-A2B

LiquidAI · on Arena since 12 Oct 2026

Download New · waiting for Arena votes
Arena score—Not voted on Arena yet
Size24 B2 B active per token (mixture of experts)
4-bit download15 GBweights only, Q4_K_M
Max context128Ktokens, from the model config

Where to use it for free

New: waiting for Arena votes

Released on 12 Oct 2026. Arena has not voted on this model yet. As soon as it does, the model moves into the ranking automatically, usually the next night. Until then you can still check its size, memory needs and speed below.

Speed on popular hardware

HardwareGPU memoryFits?Speed
RTX 3060 12 GB12 GBGPU + 64 GB RAM97 tok/s
RTX 4060 Ti 16 GB16 GBFits in GPU107 tok/s
RTX 309024 GBFits in GPU348 tok/s
RTX 409024 GBFits in GPU375 tok/s
RTX 509032 GBFits in GPU666 tok/s
RX 7900 XTX24 GBFits in GPU357 tok/s
2 × RTX 309048 GBFits in GPU338 tok/s
4 × RTX 309096 GBFits in GPU319 tok/s
Mac M4 Max 128 GB96 GBFits in GPU203 tok/s
Ryzen AI Max+ 395 128 GB96 GBFits in GPU95 tok/s
8 × H100 SXM640 GBFits in GPU1029 tok/s

Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.

GPU memory needed

Level8K context32K context128K contextQuality vs full
Q3_K_M13 GB13 GB15 GB+7.6% noticeable loss
Q4_K_M15 GB16 GB18 GB+2.0% small loss
Q5_K_M18 GB18 GB20 GB+0.7% very small loss
Q6_K20 GB21 GB23 GB+0.3% near-lossless
Q8_026 GB27 GB29 GB≈ 0% same as full
F1649 GB49 GB51 GB≈ 0% same as full

Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.