Models - Aug 5, 2026

LFM2.5-2.6B Phone and CPU Memory and Speed Guide

Q: How fast is LFM2.5-2.6B on CPU?

Liquid AI reports 220 tokens per second on an Apple M5 Max and 113 on an AMD Ryzen AI Max+ 395 in its tests. Those provider results do not predict another device, quantization, prompt length, runtime, or sustained thermal condition.

Quick answer

Liquid AI positions LFM2.5-2.6B for phones and laptops and reports strong device results. Treat them as provider-published reference points, not a device guarantee.

The model card reports 220 tokens/s on an Apple M5 Max, 113 tokens/s on an AMD Ryzen AI Max+ 395, and a 30-token/s phone result, with an under-2.5 GB memory statement attached to its launch summary. Your result can change with artifact, quantization, runtime, context, prompt processing, sampling, concurrency, thermals, and background services.

Build the memory budget

peak memory = model weights
            + KV cache at tested context
            + runtime and graph buffers
            + tokenizer and prompt workspace
            + retrieval index and tool processes
            + operating-system headroom
            + safety margin

The 131,072-token maximum context is not a promise that a constrained device can use the full window at acceptable speed. KV-cache use grows with active context, and long prompts add prefill time even when the weights fit.

Test three workload points

Test	Purpose	Record
Short smoke test	Verify artifact and chat template	Load time, first token, output correctness
Typical workflow	Represent real prompts and tool traces	p50/p95 latency, decode rate, peak memory, battery
Stress boundary	Find safe context and duration limits	OOM, throttling, heat, cancellation, recovery

Run enough repetitions to separate warm-up from steady state. On phones and fanless laptops, test sustained work rather than a short burst. Record ambient conditions and whether the runtime used CPU, GPU, NPU, or mixed execution.

Choose quantization by accepted work

GGUF and ONNX publish multiple precision or quantization options; MLX also provides Apple Silicon packaging. Smaller artifacts may reduce memory and improve throughput, but the correct choice is the smallest version that still passes the task evaluation. Measure extraction accuracy, tool arguments, refusal behavior, long-context retrieval, and final answers—not tokens per second alone.

Use the runtime guide to pick the package and the tool-harness guide to include external tool overhead and behavior.

Frequently asked questions

How much memory does LFM2.5-2.6B need?

There is no single safe number. Budget the selected weight artifact, KV cache for the actual context, runtime buffers, operating-system headroom, retrieval and tool processes, and a safety margin, then measure peak resident memory on the target.

How fast is LFM2.5-2.6B on CPU?

Liquid AI reports 220 tokens/s on an Apple M5 Max and 113 on an AMD Ryzen AI Max+ 395 in its tests. Those provider results do not predict another device, quantization, prompt length, runtime, or sustained thermal condition.

Can LFM2.5-2.6B run on a phone?

Liquid AI says the model is small enough for a phone and reports a 30-token/s phone result. Confirm your exact device, artifact, context, runtime, battery, thermals, tool stack, and accepted-task latency before making a fit claim.

Official sources

Source check: August 5, 2026. Reproduce performance and memory results on the exact target before relying on them.