Liquid AI Releases LFM2.5-2.6B, a 2.6B-Parameter Agent Model for On-Device Deployment

Liquid AI ships a 2.6B model built for local agentic workloads

Liquid AI released LFM2.5-2.6B on August 4, 2026, a 2.6-billion-parameter model designed specifically to run tool-calling agents on-device — from laptops to phones — without a cloud inference bill. The model and its base variant, LFM2.5-2.6B-Base, are available now on Hugging Face.

The model was pre-trained on approximately 34 trillion tokens, with a mid-training phase that extends its context window to 128,000 tokens. According to Liquid AI, LFM2.5-2.6B is the fastest model the company tested at its size class, decoding at 220 tokens/second on an Apple M5 Max and 113 tokens/second on an AMD Ryzen AI Max+ 395 CPU, all within under 2.5 GB of memory. The company also claims the model can run at 30 tokens/second on phone hardware.

On GPU, Liquid AI claims the model reaches nearly 15,000 output tokens per second at high concurrency, translating to roughly 1.3 billion tokens per day on a single H100.

Training pipeline: four post-training stages

Liquid AI built LFM2.5-2.6B through a four-stage post-training process on top of the pre-trained base model:

1. Supervised fine-tuning (SFT) — two rounds weighted toward agentic data including tool use, web search, and harness trajectories.
2. Teacher specialization — separate specialist teacher models trained per domain (math, code, tool use, and others).
3. Multi-domain on-policy distillation (MOPD) — specialist teachers distilled into a single student model.
4. Agentic Reinforcement Learning (Agentic RL) — multi-turn RL run inside real agent harnesses such as OpenClaw and Hermes Agent, using a system that separates model optimization, rollout generation, and sandboxed environment execution.

Benchmark claims

Liquid AI compared LFM2.5-2.6B against models up to roughly 4x larger, including Gemma-4-E2B-it (5.1B), Gemma-4-E4B-it (8B), Qwen3.5-4B (4.7B), and Qwen3.5-9B (9.7B). According to the company's own benchmarks:

• Instruction following: LFM2.5-2.6B topped every benchmark tested, including IFBench (59.17), Multi-IF (80.07), and IFStruct (85.49).
• Tool use: The model led every tool-use benchmark except BFCLv4 (56.88), where the 9.7B Qwen3.5-9B scored higher at 60.13. On ToolSandbox, LFM2.5-2.6B scored 77.83.
• Agentic tasks: On Claw-Eval average (EN), it scored 62.85, ahead of both Gemma models and close to the Qwen3.5 models (58.02–66.53 range). On BrowseComp+ (OpenClaw), it scored 26.89.
• Math: AIME25 score of 51.87, trailing Qwen3.5-9B's 56.07 but ahead of both Gemma variants.
• Coding: LiveCodeBench v6 score of 59.41, behind larger models — Liquid AI acknowledges coding is where bigger models retain a clear advantage.

These figures come directly from Liquid AI's own published benchmarks and have not been independently verified.

Deployment and access

LFM2.5-2.6B has day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX runtimes, and integrates with transformers version 5.0.0 or later. Liquid AI also published a WebGPU browser demo showing the model powering a research agent, and a guide for running it inside agent harnesses including OpenClaw, Hermes Agent, and Pi. As an open-weight release, there is no per-token API pricing; developers self-host the model.

What this means

LFM2.5-2.6B is part of a broader push toward small models specialized for agentic tool-calling rather than general knowledge or raw reasoning depth. The tradeoff is explicit in Liquid AI's own numbers: strong on instruction following and tool use, weaker on coding compared to larger competitors. For developers building on-device assistants, browser agents, or privacy-sensitive tools where sending data to a cloud API isn't viable, a model that runs under 2.5GB of memory and hits usable speeds on CPU is a meaningful option — provided the benchmark claims hold up under independent testing. The real test will be third-party evaluation on agentic harnesses beyond Liquid AI's own Claw-Eval and PinchBench suites.