LFM2.5-2.6B Release and Agent Focus
Liquid AI today released LFM2.5-2.6B, a 2.6 billion parameter agentic model designed to run entirely on-device without any cloud dependency. The company created LFM2.5-2.6B specifically for planning, tool use and multi-step tasks that previously required models three or four times its size. This approach directly responds to the persistent frustration of on-device AI: models small enough to fit on a phone often collapse when asked to handle multi-step logic or call external tools reliably.
The promise is zero marginal cost agents. Because LFM2.5-2.6B runs locally, developers can spawn parallel agent processes for calendar monitoring, travel booking, file organization or other background activities without worrying about per-token API bills. Liquid AI’s own release blog positions the model as a way to shift agent workloads onto user hardware, lowering latency and removing privacy trade-offs that come with cloud inference.
The model is already available on Hugging Face in full and quantized versions. It ships in GGUF format for llama.cpp, LM Studio and Ollama, as well as MLX and vLLM. While the release widens access to capable local agents, the company openly notes that LFM2.5-2.6B is not a replacement for larger models on heavy coding or knowledge-intensive benchmarks. That gap defines where the real work still sits.
How Liquid AI Built LFM2.5-2.6B
Building a 2.6B parameter model that punches above its weight meant rethinking the transformer’s attention-heavy blueprint. LFM2.5-2.6B uses the LFM2 hybrid architecture, which combines efficient gated short convolution blocks with a small number of grouped query attention (GQA) layers. According to an AlphaSignal breakdown, only 37 percent of the model’s layers are attention blocks. The remaining 63 percent are cheaper convolution blocks that maintain a fixed-size state, avoiding the quadratic memory growth and large KV caches found in standard transformers.
This design lets LFM2.5-2.6B handle sequences up to 128K tokens while keeping memory consumption under 2.5 GB. Liquid AI ran hardware-aware architecture search across candidates that included Mamba, Mamba2, linear attention and S4, eventually landing on short convolutions paired with a handful of GQA layers. The result is a model that can retrieve long-range context when needed through its attention blocks but does the bulk of its compute in linear-time convolution operations.
Training was equally unorthodox in scale for a tiny model. LFM2.5-2.6B was trained on 34 trillion tokens. The post-training recipe included supervised fine-tuning, knowledge distillation from multiple specialized teacher models and a final agentic reinforcement learning stage that used GRPO inside real tool harnesses such as Hermes Agent, OpenClaw and Pi. This multistage pipeline is what gives LFM2.5-2.6B its agentic edge, but the same pipeline means that gains may not transfer evenly to open-ended reasoning or coding tasks.
Benchmark Wins Over Qwen3.5-9B
Despite being roughly one-quarter the parameter count, LFM2.5-2.6B beats Qwen3.5-9B on several agentic benchmarks. Numbers shared by Liquid AI and verified by community testers on X show LFM2.5-2.6B reaching 77.83 on ToolSandbox compared to Qwen’s 76.44. On τ³ Banking, a realistic tool-calling suite, LFM2.5-2.6B scores 5.67 versus 5.15 for the 9B model. Instruction-following metrics tell a similar story: LFM2.5-2.6B hits 59.17 on IFBench (Qwen 56.47), 80.07 on Multi-IF (62.55) and 85.49 on IFStruct (78.50).
These are not blanket victories. Qwen3.5-9B still wins on several other shared benchmarks, and Liquid AI clearly states that LFM2.5-2.6B is not recommended for coding-heavy or knowledge-intensive agentic workloads. The model’s strength lies in coordinating tools, following multi-step plans and staying reliable inside agent loops, not in open-ended code generation.
The focus on instruction fidelity and tool execution means LFM2.5-2.6B could become a strong base for local agents that automate practical phone or laptop tasks. The real test, however, will be performance on messy real-world JSON payloads rather than clean test fixtures, a point raised by developers online. The next phase of deployment will show whether these benchmark numbers hold up when agents start calling live APIs.
On-Device Speed and Memory Profile
Speed is where LFM2.5-2.6B really leans into its small footprint. On an Apple M5 Max, the model runs at roughly 220 tokens per second. AMD’s Ryzen AI MAX+ 395 hits about 113 tokens per second, and a phone can manage around 30 tokens per second. All of this happens within 2.5 GB of memory, making LFM2.5-2.6B practical for background agent tasks on consumer hardware.
Those numbers come from benchmarks published on AlphaSignal and align with Liquid AI’s emphasis on low-latency inference. The model supports GGUF quants that shrink it further while preserving most of its agent capabilities. This opens the door to running multiple instances in parallel, say a desktop organizer agent alongside a travel booking loop, all without touching a remote server.
What remains unresolved is how real-world agent tasks with heavy state tracking or large prompt histories will affect throughput. The model’s 128K context window is generous, but sustained multi-turn tool use could still stress memory budgets on phones. Early adopters will need to test whether the promised 30 tok/s holds during complex chains of API calls.
Agent Tasks Where 2.6B Is Not Enough
LFM2.5-2.6B makes a clear trade-off. It excels at planning and tool orchestration but intentionally steers away from code-heavy or deep reasoning workloads. Liquid AI’s own guidance recommends larger models for those tasks, a candid carve-out that helps set expectations for developers who might otherwise treat the model as a universal miniaturized replacement.
The wider story here isn’t one model; it’s a shift toward intelligence density. Training a small model on 34 trillion tokens with a specialized agentic post-training pipeline signals that the next local AI race won’t be about sheer parameter count. LFM2.5-2.6B shows that with enough training data and a focused architecture, small models can handle real agent work. The ceiling for what these tiny models can do still sits ahead, not behind.
Frequently Asked Questions
How can I run LFM2.5-2.6B on my phone?
LFM2.5-2.6B is available in GGUF quantized formats that work with llama.cpp, Ollama and LM Studio. You can load one of the smaller quantized versions on Android or iOS devices with enough memory, targeting roughly 30 tokens per second. The Hugging Face page lists quants from 4-bit upwards, so pick one that fits your device’s RAM.
What tasks does LFM2.5-2.6B perform best?
The model is optimized for multi-step agentic tasks: calendar monitoring, travel booking, file management and tool-calling workflows where it needs to follow instructions and chain multiple API calls. It shines in benchmarks like ToolSandbox and τ³ Banking but is not built for heavy coding or long knowledge retrieval jobs.
Can LFM2.5-2.6B replace larger models for everyday coding?
No. Liquid AI explicitly advises against using LFM2.5-2.6B for coding-heavy or knowledge-intensive agentic tasks. Larger frontier models still hold a clear advantage on those benchmarks. Think of LFM2.5-2.6B as a specialist agent orchestrator rather than a general-purpose coder.
