LFM2.5-24B-A2B for On-Device Agents: What Deployment Teams Need to Verify

Liquid AI presents LFM2.5-24B-A2B as a model for fast, private agents running on-device, and reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395. ¹ Those figures are a reason to evaluate the model—not a reason to approve it for your devices, workloads, or privacy requirements.

What the LFM2.5-24B-A2B announcement establishes—and what it leaves open

LFM2.5-24B-A2B is a 24-billion-parameter mixture-of-experts model with 2 billion active parameters per token. Liquid AI says the model and its Base version are open-weight and available on Hugging Face. ¹ The architecture has 40 layers, 64 experts per MoE block with top-four routing, and a hybrid of gated short convolutions and grouped-query attention. ¹

The context window grows from 32K to 131K tokens, and the vocabulary from 65K to 128K tokens. Liquid AI reports pre-training scaling from about 17T to 33T tokens, followed by mid-training with 2T tokens at 32K context and 300B tokens at 131K context. ¹ These details describe the model's development and design; they do not establish memory use for a particular runtime configuration.

The gap between the pitch and the evidence matters. The reported decoding rates are tied to two named devices, but the source does not provide the configuration details needed to reproduce those measurements. ¹ “Private” is part of the product positioning; the source does not describe data flows, logging behavior, security controls, or threat-model assumptions. ¹ Teams facing the same gap in other local or isolated setups may find Deepgram's air-gapped voice AI push and its open deployment questions a useful parallel.

Analysis: Treat the announcement as a starting point for evaluation, not a production-readiness verdict. Parameter count, context limit, and reported benchmark results do not establish whether the model fits your device's memory budget, meets your latency target, or keeps sensitive data within the boundary you intend to define.

Local inference runtimes: which deployment paths are documented?

Liquid AI lists day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. ¹ Its Transformers example specifies Transformers 5.0.0 or later, loads the model in bfloat16 with device_map="auto", and presents Flash Attention 2 as an option on compatible GPUs. ¹

model = AutoModelForCausalLM.from_pretrained(
    "LiquidAI/LFM2.5-24B-A2B",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

For local agents, the post recommends serving the model behind an OpenAI-compatible backend. It lists ports of 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang, and 1234 for LM Studio. ¹ An API-compatible interface can simplify agent-side integration, but it does not establish that runtime behavior is interchangeable. Memory use, supported features, and latency still need to be checked for the stack you choose.

How to pilot on-device model deployment

A practical pilot should compare two listed runtimes on your target hardware—for example, MLX against llama.cpp, or vLLM against SGLang—and measure the workloads that matter to your deployment. That is a recommendation for evaluation, not a result reported by Liquid AI.

One reading is that a local model does not remove the need for controls around the agent itself. Permissions, logging, and shutdown paths still apply, which is the subject of runtime governance for always-on AI agents. If the pilot also compares hosted models against the local one, a governed model catalog can keep that comparison in one place.