Liquid AI Ships LFM2.5-24B-A2B for On-Device Agents
On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, the largest and final model in its LFM2.5 family. The model runs 24 billion parameters through a mixture-of-experts (MoE) architecture that activates only 2 billion at each inference step, producing frontier-capable reasoning at a fraction of the compute cost. It is open-weight, available on Hugging Face, and built to run entirely on-device, with no cloud API required.
For enterprise AI teams, the release closes a gap that has slowed private-agent deployments since frontier models moved to cloud-only APIs: you can now run a reasoning agent with long context, tool calling, and 24B parameters on hardware you already own.
What Changed and Why It Matters
The LFM2.5 family now spans from the 230M-parameter LFM2.5-230M to this 24B model, covering two orders of magnitude with a consistent architecture. Each tier is built for the same use case: agents that call tools, manage multi-step workflows, and run without phoning home.
Until now, enterprise teams building private agents had to choose between small models that lacked sufficient reasoning depth and large cloud-hosted models that introduced data-transfer risk and per-token costs that scaled badly at volume. LFM2.5-24B-A2B positions itself as a third option: large enough to reason reliably, fast enough to feel interactive on consumer hardware, and deployable under your own data perimeter.
Three enterprise implications stand out:
Cost at volume. Liquid AI reports that the active-parameter design decouples model size from inference cost. An agent running background scans, document classification, or signal routing can burn millions of tokens per day. At the cloud API price of comparable reasoning models, that volume breaks most per-token budgets. Running LFM2.5-24B-A2B locally eliminates marginal token cost entirely for on-device runs, and the Liquid Console hosted API price of $0.10 input and $1.00 output per million tokens is among the lowest available for a model in this capability range.
Data residency. For finance, legal, and healthcare teams, cloud transmission of sensitive documents is often non-negotiable. A model that runs on a developer’s workstation or a team’s private inference server removes the residency question entirely. The ONNX export and GGUF formats further reduce the barriers to on-premises deployment.
Parallelism. Agents that run locally can be massively parallelized across available hardware. A team with eight workstations can run eight simultaneous agent threads at no marginal API cost, opening architectural patterns that per-token pricing makes impractical.
Technical Specifications
| Attribute | LFM2.5-24B-A2B |
|---|---|
| Total parameters | 24 billion |
| Active parameters | 2 billion (MoE) |
| Context window | 131,072 tokens |
| Throughput (M5 Max) | 184 tokens/sec |
| Throughput (Ryzen AI Max+ 395) | 114 tokens/sec |
| Hosted API input price | $0.10 per million tokens |
| Hosted API output price | $1.00 per million tokens |
| License | Open-weight (see model card) |
| Available formats | Native, FP8, GGUF, ONNX, MLX-4bit |
| Speculative decoding draft | DSpark (coming soon) |
The 131K context window covers most enterprise document workloads: a typical 50-page contract, a multi-channel support thread, or several months of CRM activity notes fit comfortably in a single context. The model was pre-trained on 2 trillion tokens with an extended 1M-context phase, then post-trained through supervised fine-tuning, teacher specialization, two multi-domain on-policy distillation stages, and agentic reinforcement learning inside the harnesses it ships with.
That last step matters for reliability. The model was trained to call tools the way Hermes Agent, OpenClaw, Pi, and OpenCode actually structure their tool calls, rather than on synthetic data. Liquid AI reports this reduces dispatch errors and hallucinated tool arguments in real agent loops.
Deployment Paths
Liquid AI provides four integration paths in its documentation:
Local inference. Run the GGUF variant with llama.cpp or the MLX-4bit variant on Apple Silicon. Expose an OpenAI-compatible endpoint locally, then point any agent harness at it. No code changes are required if the harness already supports an OpenAI-style API.
High-throughput GPU serving. The FP8 checkpoint reduces GPU memory requirements and increases throughput on vLLM and SGLang. The upcoming DSpark draft model enables speculative decoding with SGLang for further speed gains.
Cross-platform edge. The ONNX export targets deployment on devices that lack GPU support, including industrial edge hardware, IoT systems, and corporate workstations without discrete graphics.
Liquid Console API. For teams that want hosted inference without managing their own server, the Liquid Console API provides the same model at $0.10 per million input tokens, with $2 in sign-up credit available. This path suits development and lower-volume production workloads.
Where This Fits in the Enterprise Agent Stack
Running capable agents on-device does not mean running them in isolation. The tool-calling architecture assumes the agent reaches external systems for retrieval, action, and verification. The practical deployment pattern is a local inference server for the model plus MCP or API connections to the enterprise systems the agent needs to act on.
For teams building the multi-agent governance architectures that are standard in mature agentic deployments, LFM2.5-24B-A2B serves well as a reasoning node inside an orchestration layer. It handles multi-step planning and tool dispatch locally, while a central orchestration layer handles routing, audit logging, and cross-agent coordination.
For teams building AI content and operational systems, the model’s speed on Apple Silicon is particularly relevant: editorial agents that classify, route, tag, and draft short-form content can run continuously on a standard developer machine or a small server, without cloud API exposure.
The release also provides a useful cost anchor for enterprise teams evaluating model strategy. The hosted API price of $1.00 per million output tokens is roughly 10 times lower than Claude Opus 5.5 and 5 times lower than GPT-6 Sol. For workflows where reasoning depth matters but absolute frontier capability is not required, LFM2.5-24B-A2B offers a defensible trade-off. Teams that have been using cloud-hosted agents APIs for production deployments should now benchmark against local LFM2.5-24B-A2B deployment for their high-volume internal workflows.
Next Steps
The base model (LFM2.5-24B-A2B-Base) is also open-weight on Hugging Face, enabling fine-tuning for domain-specific tasks. Liquid AI has not announced a timeline for the next model family, but the LFM2.5 family is described as complete with this release. The DSpark speculative decoding draft is listed as coming soon for the SGLang path.
Enterprise teams evaluating private-agent infrastructure should run a benchmark on representative workflows before the DSpark release, which is likely to add meaningful additional throughput on GPU hardware.
Book a call with Enera at /book-a-call to discuss how on-device agent models fit your enterprise deployment architecture.