The active path is not the whole shipment
Liquid AI describes LFM2.5-24B-A2B as a 24B-parameter mixture-of-experts model with 2B active parameters per token. The model uses 64 experts in each mixture block and routes each token through four of them. That active count helps explain work per token. It does not say that the other weights vanish from the model package.
The Hugging Face API lists ten safetensor shards at revision 1769516a. Those shards total 47,944,143,456 bytes, or 47.94 GB. The publisher's Q4_K_M GGUF is 14.53 GB. Its MLX 4-bit repository totals 13.59 GB. File size still is not runtime memory. The runtime can add cache, buffers, code, and operating-system pressure.
Active parameters describe the firing lane. Stored weights describe the cargo bay.
The headline speed has a test bench
Liquid AI reports 184 output tokens per second on an Apple M5 Max and more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency. The article also states the GPU setup: SGLang 0.5.22, 1,024 input tokens, up to 256 output tokens, BF16, and a sustained request load.
Those are publisher measurements on named machines. They do not predict cold start on a smaller Mac, one-user latency on a workstation, or throughput after a long agent transcript grows the key-value cache. Copy the test shape before copying the number.
The package decides the first road test
The model page exposes full-precision, GGUF, MLX, and ONNX routes. The documentation warns that MLX needs version 0.32.0 or newer because earlier versions route the experts incorrectly. That is not a tuning suggestion. It is part of the correctness floor.
Start with the artifact that matches the machine and runtime. Record its exact revision and bytes. Then measure load time, settled memory, first-token delay, sustained generation, and cache growth with your context length. A quantized file can fit on disk and still leave too little memory for the working session.
Agent fit needs an agent workload
The release reports results across tool-use, instruction-following, knowledge, math, and coding benchmarks. It leads or matches the compared models on five of eight tool-use and agent benchmarks in the publisher's table. Qwen3.6-35B-A3B leads the same table on WebArena Verified and SWE-bench Verified. The mixed result is useful. No single score settles local-agent fit.
Use your tool schemas, repository shape, stop rules, and failure costs. Test malformed calls, stale state, denied writes, long outputs, and recovery after a wrong tool choice. A model that answers quickly but repeats a bad action is not fast in the only sense that matters.
Run five separate gauges
- Pin the model and package revision. Record the exact artifact bytes.
- Measure cold start, first token, sustained output, settled memory, and cache growth.
- Replay your real tools and held-out tasks. Keep read, write, and stop cases separate.
- Log every attempted action and the result. Test recovery after the first wrong turn.
- Compare the local route with a smaller model, a rule, and the current hosted route.