On August 4, 2026, Liquid AI released LFM2.5-2.6B - a 2.6-billion-parameter model designed to run full AI agents on consumer hardware, not just answer questions.
It scores 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B, outperforming a model four times its size on one of the most predictive agentic benchmarks in the current evaluation suite. That number is worth sitting with: a model that fits on a phone beating a 9B-parameter model at the task most relevant to agent pipelines.
What makes LFM2.5-2.6B architecturally different from a transformer
Most small models trying to run on-device are just shrunken transformers. LFM2.5-2.6B is not. The hybrid LFM2 design uses 22 short convolution blocks and only 8 attention layers, trained on 34 trillion tokens with a 128K context window. That ratio matters more than the parameter count.
In a traditional transformer, every layer uses softmax attention, which scales quadratically with sequence length and produces KV caches that consume VRAM. The LFM2 convolution blocks sidestep this entirely - each convolution layer maintains a fixed-size state that gets updated on each new token, with no attention over past tokens and no KV cache growth: the state memory at token 10,000 is the same as at token 10.
To be precise about this: LFM2 does have a KV cache, from its GQA attention layers. The accurate description is that the KV cache is significantly reduced compared to a pure transformer of the same size
- not eliminated. That distinction matters if you are planning long-context agentic sessions and comparing memory projections.
Liquid AI is an MIT CSAIL spinout founded in 2022. The company name comes from "Liquid Neural Networks," inspired by the nervous system of nematodes
- systems where neurons change behavior based on input in real time. The LFM2 design is explicitly edge-first: architecture, pre-training, and post-training are co-designed around maximizing downstream quality subject to device-side latency and peak memory constraints. That is a different optimization target than frontier labs use, and it shows in the results at 2.6B parameters.
How the model was trained to do agent work, not just chat
Benchmark wins at this scale usually come from one of two places: an unusual architecture or a very targeted post-training pipeline. LFM2.5-2.6B has both.
The four-stage post-training pipeline includes agentic reinforcement learning run inside real harnesses - specifically Hermes Agent, OpenClaw, and Pi - using GRPO.
Training ran inside live agent harnesses via a Harness Proxy that captures trajectories without modifying existing frameworks, exposing the model to real tool interactions during RL.
That is meaningfully different from the standard approach of fine-tuning on synthetic trajectories or static function-calling datasets. The model learned to use tools by actually using them during training, in the same frameworks developers deploy it into. The downstream effect is that it handles multi-step tool calls and instruction sequences that trip up models trained on cleaner, curated data.
Day-one support spans llama.cpp, MLX, vLLM, SGLang, and ONNX across AMD, Qualcomm, and Apple Silicon hardware. That is a wider inference ecosystem than most new releases manage at launch, and it removes the usual friction of checking whether your stack supports the architecture.
Beagle in action#platform-eng, 11:02amWhere it actually falls short
LFM2.5-2.6B is not recommended for coding-heavy or knowledge-intensive agentic tasks; larger models still hold an edge there. Liquid AI says this directly in the model card, which is the kind of honest limitation that usually gets buried.
The benchmark selection is also telling. The comparison set Liquid AI chose - ToolSandbox, IFEval, IFBench, and CaseReportBench - are exactly the benchmarks where instruction-following and tool-routing matter most. The AIME25/LiveCodeBench gap against Qwen3.5-9B is the honest asterisk: this is a tool-use and instruction-following specialist, not a frontier reasoning model, and Liquid AI's benchmark selection reflects that focus rather than obscuring it.
For teams where the primary use case is routing tool calls, running structured data extraction, or keeping sensitive context off third-party servers, those gaps may not matter. For teams who need the model to write and debug production code, they will.
There is one more thing to check before you deploy: LFM2 was Apache 2.0-licensed, but LFM2.5 switched to the proprietary LFM Open License v1.0. That is a meaningful change if you are building a commercial product on top of it. Review the terms before you build a dependency.
| Dimension | LFM2.5-2.6B | Qwen3.5-9B |
|---|---|---|
| Parameters | 2.6B | 9.7B |
| ToolSandbox | 77.83 | 76.44 |
| Peak memory | ~2.5 GB | ~6-8 GB |
| Decode (M5 Max) | 220 tok/s | ~80-100 tok/s est. |
| Coding/reasoning | Limited | Stronger |
| License | LFM Open License v1.0 | Apache 2.0 |
| On-device RL training | Yes (real harnesses) | No |
On-device AI agent models: common questions
What is LFM2.5-2.6B?
LFM2.5-2.6B is an open-weight, 2.6-billion-parameter model from Liquid AI, released August 4, 2026. It is built on a hybrid architecture combining convolution blocks with grouped-query attention, optimized for running full agentic workflows on-device - including multi-step tool calling - at under 2.5 GB of memory.
Can LFM2.5-2.6B really beat a 9B model on agent tasks?
On ToolSandbox, yes. LFM2.5-2.6B scores 77.83 versus 76.44 for Qwen3.5-9B. The win comes from two things: a KV-cache-efficient architecture that stays fast on constrained hardware, and post-training run inside real agent harnesses rather than on static datasets. It does not beat larger models on coding or pure reasoning.
Is LFM2.5-2.6B open source?
It is open-weight: the weights are freely downloadable on Hugging Face. It is not fully open source. The training data and pipeline are private, and the model ships under the LFM Open License v1.0 - not Apache 2.0, which LFM2 used. Check the license terms before building a commercial product on it.
What hardware does LFM2.5-2.6B need?
It fits in 2.5 GB of memory and decodes at 220 tokens per second on an Apple M5 Max, 113 tokens per second on an AMD Ryzen AI Max+ 395 CPU, and around 30 tokens per second on a phone. Day-one framework support includes llama.cpp, MLX, vLLM, SGLang, and ONNX.
What is the best use case for a small on-device agent model?
Tool routing, structured data extraction, and multi-step instruction following on sensitive data that cannot leave a device. Teams handling patient records, legal documents, or internal code repositories with compliance requirements are the natural fit. For heavy code generation or deep reasoning tasks, a larger cloud model still outperforms what fits in 2.5 GB today.