Liquid AI has put out LFM2.5-24B-A2B, an open-weight Mixture-of-Experts reasoning model with 24B total parameters and roughly 2B active per token. The pitch is explicit: agents that run on the device in front of you. The numbers attached are 48.09 on LiveClaw and 184 tokens per second decoding on an Apple M5 Max.
That throughput figure is the one worth sitting with. Agent loops are not chat. A single useful agent task means tool calls, failed tool calls, re-reads, retries, and a lot of tokens the user never sees. At 20 tokens/s a local agent is a demo. At 184 it is something you can leave running while you do other work.
Sparse is the whole trick
The architecture is doing the heavy lifting here, and it’s the same bet a lot of labs are now making at smaller scale. You hold all 24B weights in memory, but each token only wakes up about 2B of them. Memory cost scales with the full model; compute cost scales with the active slice. On a machine with plenty of unified memory and a modest amount of compute per watt, that trade is close to ideal. It’s a very deliberate shape for Apple silicon and for the laptop-class hardware most people actually own.
The thing dense small models keep running into is that compressing capability into 4B or 8B parameters costs you exactly the skills agents need most: instruction adherence across long tool chains, and not forgetting what you were doing six steps ago. Sparse MoE sidesteps some of that by keeping more total knowledge around. Whether 2B active is enough to steer it reliably is the open question, and one benchmark score does not answer it.
A number without a scoreboard
48.09 on LiveClaw is published without much company. I’d like to see it next to the obvious comparisons at similar active-parameter budgets, and next to the frontier APIs people would otherwise call. Agentic benchmarks are also young and noisy in a way that MMLU-style evals are not; scaffolding, retry policy and tool definitions move scores more than most write-ups admit. Treat the figure as a signal that the model is in the conversation, not as a ranking.
What makes this release matter more than the score is the privacy story being structural rather than promised. An agent that reads your files, your mail, your repo and your calendar is the single most invasive piece of software you will ever install. Every remote version of that ships your context to someone else’s logs. A local one does not, and no terms of service update can change that.
Open weights, so this gets tested properly within days by people who did not write the benchmark. The interesting experiment is not the M5 Max, which is a well-provisioned machine. It’s what 184 tokens/s becomes on a two-year-old laptop with 16GB of RAM.