Can I run LFM2.5 8B A1B (MoE) locally?
LFM2.5 8B A1B (MoE) has 8.47 billion parameters (Hugging Face). At the popular Q4_K_M quantization it needs roughly 6.0 GB of GPU memory ESTIMATE. It has been on Hugging Face since 2026-05-28. It's a mixture-of-experts model: all weights must fit in memory, but it generates faster than a dense model of the same size.
Memory needed by quantization
| Quant | Weights | Total needed (est.) |
|---|---|---|
| IQ1_S | 1.7 GB | 2.4 GB |
| IQ2_XXS | 2.2 GB | 2.9 GB |
| IQ2_M | 2.9 GB | 3.7 GB |
| Q2_K | 3.4 GB | 4.2 GB |
| IQ3_XXS | 3.2 GB | 4.1 GB |
| Q3_K_S | 3.9 GB | 4.7 GB |
| IQ3_M | 3.9 GB | 4.7 GB |
| Q3_K_M | 4.2 GB | 5.1 GB |
| IQ4_XS | 4.5 GB | 5.4 GB |
| Q4_0 | 4.8 GB | 5.7 GB |
| Q4_K_M | 5.1 GB | 6.0 GB |
| MXFP4 | 4.9 GB | 5.8 GB |
| Q5_K_M | 6.0 GB | 7.0 GB |
| Q6_K | 6.9 GB | 7.9 GB |
| Q8_0 | 9.0 GB | 10.1 GB |
| F16 | 16.9 GB | 18.4 GB |
KV cache by context length ESTIMATE
| Context | f16 | q8_0 | q4_0 |
|---|---|---|---|
| 4k | 0.05 GB | 0.03 GB | 0.01 GB |
| 8k | 0.10 GB | 0.05 GB | 0.03 GB |
| 16k | 0.20 GB | 0.11 GB | 0.06 GB |
| 32k | 0.40 GB | 0.21 GB | 0.11 GB |
| 64k | 0.81 GB | 0.43 GB | 0.23 GB |
From config.json: 6 KV-caching attention layers, 8 KV heads × 64 head dim. Max context 125k. The tables above assume 4k context.
Which GPUs and Macs can run it?
✅ fits · ⚠️ fits with little headroom · ❌ doesn't fit on this one device (multiple GPUs or RAM offload may work; check your setup). Apple Silicon assumes macOS lets the GPU use ~67–75% of unified memory. All fit verdicts are estimates.
Get the weekly "new models & what it takes to run them" email
Free. One email a week. Unsubscribe anytime.