LFM3 1.7B VRAM requirements at Q4 and 256K context

How this is calculated

The model has 24 blocks, and only some of them keep a token cache. Thirteen are convolution blocks with no token cache, six are full attention blocks and five are attention blocks with a 256 token window. The calculator's layer count is 11 because it counts the attention blocks. Hidden size is 2048, with 4 KV heads at head dimension 128. The six full blocks cost 12,298 bytes per token in FP16, and the five windowed blocks hold a fixed buffer. The convolution blocks keep a small state of about 25 KB per layer, which this estimate leaves out. Here's how the total grows: 1.18 GB at 8K, 1.51 GB at 32K, 2.84 GB at 128K and 4.61 GB at 256K. At the full context the cache is 3.3 times the size of the weights, and that's what makes context the thing to budget. At 8K the whole thing stays under 2 GB, so it fits almost anywhere. Weights are the one part you can trade for more room. Q8_0 weights come to 1.83 GB and FP16 weights to 3.46 GB, and a Q8 cache cuts the 256K total to 2.84 GB. Those are planning numbers for one sequence. Prompt processing buffers, the engine's own memory and any other apps on the device all sit on top.

Verdict

Pick a context limit before you pick a device. The full 256K scenario at 4.6 GB is over what a 4 GB memory budget can offer. At 32K the estimate of 1.5 GB leaves room on a small GPU or a phone that can give the model a couple of gigabytes. 128K at 2.8 GB still fits inside a 4 GB budget on paper. Memory is only half of the picture on a phone or laptop, because heat, battery and runtime support also set what's practical. This page doesn't predict speed or quality. Measure peak memory with your actual prompts, and check that your engine supports the conv and windowed attention blocks before you rely on the estimate.