Advisor / Models / LFM2.5 24B-A2B

LFM2.5 24B-A2B: VRAM requirements & speed

LFM2.5 24B-A2B (23.97B MoE, 2.45B active, lfm1.0). Native context 128k. Smallest GGUF: 14.53 GB.

14.53 GBsmallest quant file
23hardware pages that fit it fully
4run with offload
0measured data points

Memory needed by quantization and context

GGUF file size plus f16 KV cache (and fixed overhead of ~1 GB per GPU not shown). Source: LiquidAI/LFM2.5-24B-A2B-GGUF.

QuantFile+8k ctx+32k ctx+128k ctx
Q4_K_M14.53 GB14.7 GB15.2 GB17.2 GB
Q5_K_M17.03 GB17.2 GB17.7 GB19.7 GB
Q6_K19.69 GB19.9 GB20.4 GB22.4 GB
Q8_025.49 GB25.7 GB26.2 GB28.2 GB

KV cache size by context length

Computed from the model's attention layout.

Contextf16q8_0q4_0
8k0.17 GB0.09 GB0.05 GB
32k0.67 GB0.36 GB0.19 GB
128k2.69 GB1.43 GB0.76 GB

Which hardware runs LFM2.5 24B-A2B? · 16k context

HardwareQuantFitMemoryTok/sSpeed source
RTX 5090 32GBQ8_0fits fully25.8 GB343scaled from measured
RTX 5070 Ti 16GBQ4_K_Mfits fully14.9 GB266scaled from measured
RTX 4090 24GBQ6_Kfits fully20.0 GB265scaled from measured
RTX 5080 16GBQ4_K_Mfits fully14.9 GB263scaled from measured
AMD RX 7900 XTX 24GBQ6_Kfits fully20.0 GB228estimated
Mac Studio M5 Ultra 256GBQ8_0fits fully25.8 GB225estimated
Mac Studio M5 Ultra 512GB (ships late Oct 2026)Q8_0fits fully25.8 GB225estimated
RTX PRO 6000 96GBQ8_0fits fully25.8 GB220scaled from measured
RTX 3090 24GBQ6_Kfits fully20.0 GB206scaled from measured
2x RTX 3090 (48GB total)Q8_0fits fully25.8 GB194estimated
RTX 5060 Ti 16GBQ4_K_Mfits fully14.9 GB180scaled from measured
16GB VRAM class (RTX 4060 Ti 16GB, 32GB RAM)Q4_K_Mfits fully14.9 GB140scaled from measured
MacBook Pro 14/16 (2024) M4 Max 128GBQ8_0fits fully25.8 GB124scaled from measured
RTX 5070 12GBQ4_K_MMoE offload to CPU14.9 GB105estimated
RTX 4070 12GBQ4_K_MMoE offload to CPU14.9 GB100estimated
Mac Studio (2025) M3 Ultra 96GBQ8_0fits fully25.8 GB92scaled from measured
Mac Studio M5 Max 128GBQ8_0fits fully25.8 GB87scaled from measured
MacBook Pro 14/16 M5 Max 128GBQ8_0fits fully25.8 GB87scaled from measured
AMD Ryzen AI Max+ PRO 495 (Ryzen AI Halo) 192GBQ8_0fits fully25.8 GB66estimated
NVIDIA DGX Spark 128GBQ8_0fits fully25.8 GB66scaled from measured
MacBook Pro 14/16 M5 Pro 48GBQ8_0fits fully25.8 GB65estimated
RTX 3060 12GB (32GB RAM)Q4_K_MMoE offload to CPU14.9 GB60estimated
8GB VRAM class (RTX 4060 8GB, 32GB RAM)Q4_K_MMoE offload to CPU14.9 GB55estimated
AMD Ryzen AI Max+ 395 (Strix Halo) 128GBQ8_0fits fully25.8 GB54scaled from measured
Mac mini (2024) M4 Pro 64GBQ8_0fits fully25.8 GB48estimated
Mac mini M6 32GBQ6_Kfits fully20.0 GB41estimated
MacBook Air M5 32GBQ6_Kfits fully20.0 GB39estimated
Mac mini / iMac M4 16GB-does not fit--

Discrete GPUs are modelled with 32 GB of system RAM available for offload. Does not fit even with offload: Mac mini / iMac M4 16GB.

MLX conversions for Apple Silicon

No MLX conversion yet: we found no verified mlx-community 4-bit or 8-bit repo for LFM2.5 24B-A2B. Use the GGUF files with llama.cpp or ollama on a Mac. How we check.

Links

Model card · GGUF files · Open in the advisor · last verified 2026-10-20

FAQ

How much VRAM does LFM2.5 24B-A2B need?

The smallest quant (Q4_K_M) is 14.53 GB, so you need roughly 16 GB for an 8k context including KV cache and overhead. At 128k context the f16 KV cache alone is 2.7 GB.

Can I run LFM2.5 24B-A2B on a RTX 5090 32GB?

Yes: on RTX 5090 32GB it fits fully as Q8_0 at about 343 tok/s (scaled from measured). 23 of 28 hardware configurations fit it fully at 16k context.

What is the maximum context for LFM2.5 24B-A2B?

128k tokens natively. Memory, not the model, is usually the limit: see the KV cache table above.

All models: model index · by memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB. Numbers use the same fit math as the advisor; sources and methods.