# Liquid AI Releases LFM2.5-2.6B On-Device Agentic Model

> The open-weight release brings planning, tool calling and multi-step agent workflows to phones, laptops and CPUs with full data locality and near-zero marginal cost while matching larger cloud models on key benchmarks.

*Published 2026-08-06 · By Diane Okafor*

LFM2.5-2.6B is a 2.6 billion parameter open-weight model optimized for fully on-device agentic workloads including planning, tool calling, and multi-step tasks.

Liquid AI has introduced the LFM2.5-2.6B model as an open-weight solution tailored for agentic operations that execute entirely on local hardware. The 2.6 billion parameter system supports planning sequences, function calls to external tools, and execution of chained tasks without any requirement for cloud connectivity. Both the base checkpoint LFM2.5-2.6B-Base and the post-trained variant are accessible immediately through Hugging Face under an open license that permits commercial use. This availability extends to immediate compatibility with llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes, lowering the barrier for integration into existing device software stacks. The design emphasizes data sovereignty because all computation occurs locally, eliminating transmission risks associated with remote inference services. Pre-training consumed approximately 34 trillion tokens, equipping the model with broad language capabilities across 16 languages while embedding native function-calling syntax. Business stakeholders evaluating on-device AI now have a concrete option that delivers performance parity with substantially larger cloud models on instruction adherence and tool orchestration tasks.

## What background led to the development of on-device agentic models like LFM2.5-2.6B?

Enterprises have increasingly sought alternatives to cloud-dependent AI because of rising concerns over data exposure, latency variability, and recurring inference expenses. Regulatory frameworks in multiple jurisdictions now require stricter controls on personal and proprietary information, pushing organizations toward architectures where sensitive context remains resident on the endpoint. At the same time, hardware improvements in mobile and laptop processors have created sufficient compute capacity for smaller yet capable models to handle structured reasoning loops. Liquid AI positioned the LFM2.5 family to exploit these hardware trends by focusing pre-training and post-training on agentic patterns rather than broad generative fluency alone. The resulting model therefore prioritizes efficient planning and tool-use circuits that can be invoked repeatedly at negligible incremental cost once the weights reside on the device. This approach contrasts with earlier generations of agent frameworks that required persistent network access and incurred per-token fees even for simple multi-step processes.

Market data from multiple analyst reports indicate that organizations deploying AI agents face cumulative costs that scale linearly with usage volume when relying on hosted endpoints. By shifting execution to the edge, the marginal cost per additional workflow invocation drops to essentially zero after initial deployment. The 128K context window further enables retention of extended conversation histories or document contexts without truncation, a capability previously limited to larger cloud instances. These factors collectively address both technical and economic constraints that have slowed enterprise adoption of autonomous agent systems in regulated industries such as healthcare, finance, and manufacturing.

## What new capabilities does the LFM2.5-2.6B model deliver for on-device agents?

The model introduces native support for structured function calling that allows it to invoke device-local or permitted external APIs within a single forward pass. This capability enables agents to decompose high-level goals into sequenced actions, select appropriate tools, and iterate based on intermediate results without external orchestration layers. The post-trained version incorporates reinforcement learning signals that improve adherence to multi-turn instructions compared with the base checkpoint. Developers can therefore deploy agents that maintain state across extended interactions while remaining fully contained within the device memory envelope of under 2.5 GB. The open license further permits fine-tuning on domain-specific tool sets, extending utility beyond the initial training distribution.

Performance characteristics include sustained throughput that reaches nearly 15,000 output tokens per second under high concurrency on a single H100 GPU according to Hugging Face benchmarks, translating to roughly 1.3 billion tokens per day of continuous operation. On CPU-only hardware the model sustains 113 tokens per second on an AMD Ryzen processor, demonstrating viability across a spectrum of deployment targets from high-end laptops to embedded systems. These figures position the 2.6 billion parameter model as a practical replacement for certain cloud workloads where privacy or cost sensitivity outweighs the need for the absolute highest benchmark scores in coding or mathematics.

## How does LFM2.5-2.6B perform on benchmarks relative to larger models?

Benchmark evaluations demonstrate that LFM2.5-2.6B matches or exceeds the instruction-following accuracy of models nearly four times its parameter count on agent-specific tasks. Tool-use success rates remain competitive because the training regimen emphasized correct sequencing and argument formatting over open-ended generation. Some coding and mathematical reasoning benchmarks show trailing performance, reflecting the deliberate trade-off of allocating capacity toward planning circuits rather than broad knowledge recall. The 128K context window contributes to sustained coherence across long-horizon workflows that would otherwise require multiple shorter calls to smaller-context models.

Technical specifications of the LFM2.5-2.6B model as reported by Liquid AI and Hugging FaceSpecificationValueParameter Count2.6 billion (2.69B total)Context Window128K tokensLanguages Supported16Inference Speed (M5 Max)220 tokens per secondMemory FootprintUnder 2.5 GBPre-training Tokens~34 trillionSupported Runtimesllama.cpp, MLX, vLLM, SGLang, ONNX

## What technical architecture supports the model's on-device efficiency?

The hybrid architecture within the LFM2.5 family combines elements of transformer attention with liquid neural network components that reduce the computational overhead of long-sequence processing. This design choice enables the observed inference speeds while maintaining the 128K context capacity. Quantization-aware training further compresses the weights to fit within the 2.5 GB memory budget on consumer silicon without significant degradation in tool-calling precision. The immediate availability of optimized kernels for Apple MLX and AMD CPU paths ensures that developers can achieve the reported throughput without custom engineering.

Deployment pipelines benefit from the model's compatibility with established inference servers, allowing seamless scaling from single-device testing to fleet-wide rollout. The open-weight release also facilitates community-driven optimizations such as additional quantization levels or hardware-specific kernels that can further improve speed on lower-end CPUs. These engineering decisions collectively lower the total cost of ownership for organizations that previously maintained separate cloud budgets for agent experimentation.

## What deployment steps enable local integration of the model?

- Download the base or post-trained checkpoint directly from the Hugging Face repository maintained by LiquidAI.
- Select and install one of the supported inference engines such as llama.cpp or MLX based on target hardware.
- Load the model weights into the chosen runtime while verifying memory allocation stays below the 2.5 GB threshold.
- Configure native function-calling templates to map model outputs to device-available tools or permitted APIs.
- Validate multi-step workflows through iterative testing on representative instruction sets before production deployment.

## What market and stakeholder implications arise from this release?

Enterprise technology leaders now possess a concrete pathway to reduce dependence on usage-based cloud billing for agentic workloads. The zero marginal cost structure favors high-volume internal automation scenarios such as compliance monitoring, internal knowledge retrieval, and routine operational workflows. Privacy officers gain assurance that proprietary data remains within organizational boundaries, satisfying data-residency requirements without additional network controls. Device manufacturers may incorporate the model into next-generation hardware bundles, creating differentiation through pre-installed agent capabilities.

Smaller development teams benefit from the open license that removes licensing friction for commercial products. The competitive benchmark results on tool use suggest that organizations can prototype agent systems locally before committing to larger hosted models for the most demanding tasks. This hybrid strategy balances performance, cost, and compliance considerations across diverse industry verticals.

## What expert reactions have been recorded regarding the LFM2.5-2.6B announcement?

> Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal cost of each run is essentially zero.Liquid AI

Industry observers note that the combination of open weights, competitive tool-use performance, and aggressive on-device optimization marks a maturation point for edge agent technology. The explicit focus on zero marginal cost addresses a longstanding objection to scaling agent deployments beyond pilot stages. While some benchmarks reveal gaps in coding and mathematics, the targeted optimization for planning and tool orchestration aligns with the primary requirements of many enterprise agent use cases.

## What developments are expected next for on-device agentic models?

Subsequent releases in the LFM2.5 series are anticipated to incorporate additional post-training stages that close remaining gaps on coding and mathematical reasoning without increasing the parameter count or memory footprint. Community contributors are expected to release fine-tuned variants for specific verticals such as legal document analysis or supply-chain optimization. Hardware vendors may announce dedicated accelerators that further increase tokens-per-second throughput on mid-range devices.

The open license encourages ecosystem growth around standardized tool-calling schemas, potentially leading to interoperable agent frameworks that function across multiple model families. Organizations that adopt the current release early can influence these standards through feedback on real-world multi-step workflow performance. Continued emphasis on data locality is likely to drive similar releases from other model developers seeking to address the same enterprise constraints.

## Sources

1. [The base and post-trained models are available on Hugging Face, with the model running entirely on-device for planning and tool calling at 220 tokens per second on M5 Max under 2.5 GB memory.](https://www.liquid.ai/blog/lfm2-5-2-6b)
2. [The model offers efficient inference with 220 tok/s on Apple M5 Max in under 2.5 GB memory and supports high concurrency throughput of almost 15K output tokens per second on a single H100.](https://huggingface.co/LiquidAI/LFM2.5-2.6B)
3. [Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal…](https://x.com/liquidai/status/2084640701669613906)

---
Source: https://aiintelreport.com/ai-agents/liquid-ai-releases-lfm2-5-2-6b-on-device-agentic-model
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt