LFM2-24B-A2B
Where it runs 8K context · ≥ 5 t/s floor
| Setup | Best plan | Memory | Generate |
|---|---|---|---|
| 8 GB GPU + 32 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 21 expert layers on CPU | Fits* | ~13–21 t/s Estimated extrapolated |
| 12 GB GPU + 32 GB RAM | Q4_K_Mllama.cpp CUDAGPU + CPU experts · 10 expert layers on CPU | Fits* | ~15–24 t/s Estimated extrapolated |
| 16 GB GPU + 64 GB RAM | Q4_K_Mllama.cpp CUDAFull GPU | Fits* | ~114–328 t/s Estimated extrapolated |
| 24 GB GPU + 64 GB RAM | Q5_K_Mllama.cpp CUDAFull GPU | Fits* | ~110–317 t/s Estimated extrapolated |
| 32 GB GPU + 128 GB RAM | Q8_0llama.cpp CUDAFull GPU | Fits* | ~132–383 t/s Estimated extrapolated |
| 2× 24 GB GPUs + 128 GB RAM | FP16llama.cpp CUDALayer split across GPUs | Fits* | ~37–108 t/s Estimated extrapolated |
| 96 GB workstation GPU + 256 GB RAM | FP16llama.cpp CUDAFull GPU | Fits* | ~71–206 t/s Estimated extrapolated |
| Mac, 24 GB | Q5_K_Mllama.cpp MetalUnified memory | Fits* | ~30–86 t/s Estimated extrapolated |
| Mac, 48 GB | MLX 8-bitMLX-LMUnified memory | Fits* | ~30–61 t/s Estimated |
| Mac, 128 GB | FP16llama.cpp MetalUnified memory | Fits* | ~22–63 t/s Estimated extrapolated |
| Mac, 512 GB | FP16llama.cpp MetalUnified memory | Fits* | ~33–94 t/s Estimated extrapolated |
| DGX Spark 128 GB | FP16llama.cpp CUDAUnified memory | Fits* | ~17–47 t/s Estimated |
| Ryzen AI Max+ 395, 128 GB | FP16llama.cpp VulkanUnified memory | Fits* | ~10–29 t/s Estimated extrapolated |
| CPU only, 64 GB DDR5-5600 dual channel | Q4_K_Mllama.cpp CPUCPU only | Fits* | ~14–16 t/s Estimated extrapolated |
Same plan engine as the planner, with its default reserves and RAM. How plans are chosen.
Files exact Hugging Face bytes
| Artifact | Size | Files | Revision | Source |
|---|---|---|---|---|
| FP16 LiquidAI/LFM2-24B-A2B-GGUF | 44.42 GiB | 1 | 6247d03b87 | HF files |
| Q8_0 LiquidAI/LFM2-24B-A2B-GGUF | 23.61 GiB | 1 | 6247d03b87 | HF files |
| Q5_K_M LiquidAI/LFM2-24B-A2B-GGUF | 15.76 GiB | 1 | 6247d03b87 | HF files |
| Q4_K_M LiquidAI/LFM2-24B-A2B-GGUF | 13.43 GiB | 1 | 6247d03b87 | HF files |
| MLX 8-bit 8-bit LiquidAI/LFM2-24B-A2B-MLX-8bit | 23.59 GiB | 5 | 387dd98f01 | HF files |
| MLX 4-bit 4-bit LiquidAI/LFM2-24B-A2B-MLX-4bit | 12.49 GiB | 3 | fb67c8c23d | HF files |
| Native safetensors mixed/custom native; ; stored tensor dtypes=F32/BF16 · F32/B LiquidAI/LFM2-24B-A2B | 44.41 GiB | 1 | a3bbacd91a | HF files |
Not found in the representative quant search (not proven unavailable): EXL3, AWQ, GPTQ. Complete selected file sets; projectors, MTP/draft and lookup files are listed separately and budgeted only by plans that use them. File bytes are not allocated memory.
Quality
| AA evaluated mode | Index | LiveCodeBench | Terminal-Bench 2.1 | GPQA | HLE | Source |
|---|---|---|---|---|---|---|
| LFM2 24B A2B AA estimate | 5.9 | 0.0% | — | 47.4% | 4.2% | source |
Arena text rating
Unknown
Publisher-reported benchmarks
None extracted from the model card.
AA and Arena describe the hosted model in the stated mode, not each quantized file; scores are never inherited by base, pruned, abliterated or quantized variants and never averaged together.
How to run it
No model-specific recipe in the research. The planner binds the backend’s documented command template (llama.cpp, Ollama, LM Studio, MLX, vLLM) to your plan: open a plan’s details.
Recipes are source examples (settings are the publisher’s or author’s, not measured optima) and were not executed by LocalFit. A downloadable file does not prove your backend build supports the architecture.
Reported speeds 0
No reported measurement yet. Submit one.