Markdown edition

# Liquid AI's LFM2.5 24B model gives local agent builders more serving choices

> Liquid AI's new LFM2.5-24B-A2B has downloadable checkpoints for local and GPU serving. Its license and separate edge and server benchmarks matter when evaluating it for tool-using agents.

By BIG CHANGE Editorial

Published: 2026-10-14T14:58:18.169Z
Updated: 2026-10-14T14:58:18.169Z
Canonical: https://bigchange.ai/blog/liquid-lfm25-24b-local-agents-serving

![Conceptual charcoal illustration of two removable storage modules beside an open compact workstation, a separate rackmount server and a small input pad on a workbench.](https://bigchange.ai/api/media/file/liquid-local-serving-components-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE; no Liquid AI hardware, agent test or benchmark result is depicted.

Liquid AI released [LFM2.5-24B-A2B](https://www.liquid.ai/blog/lfm2-5-24b-a2b) on October 13 as downloadable weights for a text-only reasoning model trained for tool use. Its [model card](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) lists checkpoints for several serving systems, from a quantized local file to formats aimed at GPU servers. For teams building an agent, those files are only one part of the setup: they still need a runtime to serve the model and a harness to execute its proposed tool calls.

## The big change

- **What changed:** Liquid AI has put the post-trained model and its base version on Hugging Face, with serving formats for different hardware. The model card specifies 24.0 billion parameters in total and 2.45 billion active during inference; the name's “A2B” is a rounded label.
- **Why it matters:** A team evaluating a local agent can choose a checkpoint and serving path before connecting a tool-using harness, rather than treating a model download as a complete agent. The license and available hardware still shape that decision.
- **What to watch:** Liquid AI reports promising speed and tool-use results under specified tests. Teams need to measure their own tasks and concurrency; the company's edge decode rate and server throughput describe different workloads.

## What a team can download

The [post-trained checkpoint](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) is intended for agentic work. A separate [base checkpoint](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-Base) is for adaptation. Liquid AI describes the new model as reasoning-only: it generates a thinking segment before its final answer, unlike the earlier LFM2-24B-A2B. Its stated context limit is 131,072 tokens, up from 32,768 in that predecessor. These are model specifications, not a measure of how well a particular application uses long context.

The model card links the native checkpoint for Transformers, vLLM and SGLang; [FP8 weights](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-FP8) for GPU serving; [GGUF files](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF) for llama.cpp; [ONNX files](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-ONNX) for ONNX Runtime; and a [4-bit MLX version](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-MLX-4bit) for Apple Silicon. BIG CHANGE confirmed those six repositories, including the base model, were accessible on October 14. The card marks a DSpark draft model for speculative decoding as “coming soon.” Its published formats should not be read as proof that every backend and hardware combination has been tested by an independent party.

The weights are **open-weight**, under the [LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/LICENSE). That license permits downloading and modification, but conditions commercial use on the user or legal entity having annual revenue below $10 million. Commercial use above that threshold is not licensed by those terms. Organizations should read the full license before deployment; “open-weight” here does not mean unrestricted commercial use.

## The model, server and harness have different jobs

The [model card's tool-use description](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) gives the model a list of functions, lets it propose a call, passes the call to software for execution, then returns the tool result so the model can answer. The model does not itself execute a browser action or file operation. The serving backend loads its checkpoint and exposes inference. An agent harness manages the exchange with tools and decides how to handle permissions and results.

Liquid AI's [local-agent instructions](https://docs.liquid.ai/examples/agent-harnesses) show that division in practice: serve the model through an OpenAI-compatible endpoint and point Hermes Agent, OpenClaw or Pi at its local address. Port and configuration depend on the selected server. The company says it trained the post-trained model in agent harnesses, including Hermes Agent and OpenCode. That is a description of training and supported setup, not proof that a downloaded agent will be reliable or that its data stays on-device in every configuration. The deployment, tools and network connections determine that boundary.

## Read the speed numbers by workload

Liquid AI reports **184 output tokens per second on an Apple M5 Max** and **114 on an AMD Ryzen AI Max+ 395** in edge decoding. The card says those tests used llama.cpp, Q4\_K\_M quantization and a 4,000-token input context, collected on Liquid AI's Pipette infrastructure. These are company measurements for a quantized local setup. They do not establish a speed for the native BF16 weights, an arbitrary laptop or a whole agent task with tools.

The company's **more than 21,000 output tokens per second** figure is aggregate GPU throughput, calculated as total output tokens divided by wall time on a single NVIDIA B300 under sustained high concurrency. Its stated setup used SGLang 0.5.22, BF16 weights, 1,024 input tokens and up to 256 output tokens, continuously replacing completed requests to maintain the target number in flight. That figure is not the response rate for one person waiting on one agent. The two measurements answer different engineering questions: local generation pace and total server capacity under load.

Liquid AI also publishes a 16-benchmark comparison against other models. For example, it reports 48.09 for LFM2.5 on LiveClaw against 38.19 for Qwen3.6-35B-A3B, while Qwen leads on WebArena Verified and SWE-Bench Verified in the same table. The company says all models were evaluated with vLLM and benchmark-specific generation settings. BIG CHANGE did not reproduce those tests, so they are useful leads for workload testing rather than an independent capability ranking.

The practical evaluation is therefore specific: pick the artifact and runtime that match the target hardware, read the license, connect only the tools the agent needs, and test the actual task and expected request load. Liquid AI offers the pieces for that evaluation now. Their behavior in a production workflow remains for each team to verify.

## Sources & further reading

- [Liquid AI, “LFM2.5-24B-A2B: Built for Fast On-Device Agents”](https://www.liquid.ai/blog/lfm2-5-24b-a2b) (October 13, 2026). The release describes what changed from LFM2, the training approach, supported serving systems and vendor-run quality and speed measurements. BIG CHANGE did not repeat its benchmarks or run the model.
- [Liquid AI, LFM2.5-24B-A2B model card](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) (read October 14, 2026). Lists 24.0B total and 2.45B active parameters, artifact links, inference settings, tool-use format, benchmark conditions and the forthcoming DSpark draft model. The card is maintained by the vendor and may change.
- [Liquid AI, LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/LICENSE) (read October 14, 2026). Sets distribution obligations and the $10 million annual-revenue condition for commercial use. Readers should use the license text for their own deployment decision.
- [Liquid AI, Agent Harnesses guide](https://docs.liquid.ai/examples/agent-harnesses) (linked from the model card, read October 14, 2026). Explains the local OpenAI-compatible serving endpoint and harness configuration. It is documentation; BIG CHANGE did not execute these instructions.

## Sources

- [Liquid AI, LFM2.5-24B-A2B: Built for Fast On-Device Agents](https://www.liquid.ai/blog/lfm2-5-24b-a2b) — Vendor release describing the model, serving ecosystem and benchmark results. The speed and quality numbers are Liquid AI measurements, not BIG CHANGE tests.
- [Liquid AI, LFM2.5-24B-A2B model card](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) — Maintained card with parameter counts, artifact links, inference and tool-use details, evaluation settings and DSpark forthcoming status. Repository SHA checked: 1769516ac588c4c558089c94f27d96b0dc67b183.
- [Liquid AI, LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/LICENSE) — The actual model license. Section 5 conditions commercial use on annual revenue below the $10 million threshold; redistribution carries notice obligations.
- [Liquid AI, Agent Harnesses guide](https://docs.liquid.ai/examples/agent-harnesses) — Maintained setup guide linked by the model card; documents local OpenAI-compatible serving and harness configuration. Instructions were not run by BIG CHANGE.