Liquid AI Launches LFM2.5-2.6B for Fast, Private On-Device AI Agents
Liquid AI launches LFM2.5-2.6B, an on-device agent model with fast CPU inference, 128K context, strong tool use, private operation and no mandatory cloud bill.
Summary
Liquid AI announced LFM2.5-2.6B on August 4, 2026, positioning the 2.6-billion-parameter model as a foundation for AI agents that run locally on phones, laptops and other edge hardware. The company says the model can call tools and complete multistep workflows without sending data to a cloud inference service. Both the agent-tuned model and LFM2.5-2.6B-Base were made available through Hugging Face, alongside a browser-based WebGPU research-agent demonstration. These availability details are directly stated in the release; performance and quality comparisons are Liquid AI’s own reported results rather than independent findings.
The model was pretrained on approximately 34 trillion tokens and received a mid-training phase that expanded its context window to 128,000 tokens. Post-training had four stages: two rounds of supervised fine-tuning weighted toward tool use, web search and agent trajectories; separate specialist teachers for areas such as mathematics, coding and tools; multi-domain on-policy distillation into one student model; and reinforcement learning conducted in multistep agent environments. Liquid AI says its training architecture separates model optimization, rollout generation and sandboxed environment execution. A proxy layer records token-level trajectories while allowing harnesses including OpenClaw and Hermes Agent to operate without modification.
In Liquid AI’s benchmark table, LFM2.5-2.6B generally performed strongly against Gemma and Qwen models ranging from 4.7 billion to 9.7 billion parameters. It recorded 85.49 on IFStruct, 80.07 on Multi-IF and 59.17 on IFBench, leading all listed competitors on those instruction-following tests. Its ToolSandbox score was 77.83, above Qwen3.5-9B’s 76.44, while its 56.88 result on BFCLv4 trailed that larger Qwen model’s 60.13. It also scored 26.89 on BrowseComp+ using OpenClaw, close to Qwen3.5-9B at 27.23. Coding remained a weaker area: its 59.41 LiveCodeBench v6 result was below Gemma-4-E4B-it at 63.77 and Qwen3.5-9B at 69.86.
Liquid AI reports decode speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, with memory use below 2.5GB. It further says the model can sustain 30 tokens per second on a phone, although the announcement does not identify the handset or configuration used. On an Nvidia H100, the company reports almost 15,000 output tokens per second at high concurrency, equivalent to about 1.3 billion tokens per day. Day-one inference support is listed for llama.cpp, MLX, vLLM, SGLang and ONNX, while Transformers 5.0.0 or newer can load the Hugging Face model directly.
The release matters to developers building private assistants, offline automation or high-volume agent services because a capable small model could reduce latency, data exposure and recurring cloud-inference costs. Its compatibility with established runtimes and agent harnesses may also lower deployment barriers. However, the article does not provide full benchmark protocols, quantization settings, energy measurements, licensing analysis or independent validation, making direct reproduction important. The next test will be whether developers can reproduce the claimed speed, memory footprint and agent reliability across ordinary phones and laptops, particularly during long-context and prolonged multistep tasks.
Positives
- LFM2.5-2.6B led every listed instruction-following benchmark, including scores of 85.49 on IFStruct, 80.07 on Multi-IF and 59.17 on IFBench.
- The 2.6-billion-parameter model scored 77.83 on ToolSandbox, exceeding all four larger Gemma and Qwen models in Liquid AI’s comparison.
- Liquid AI reports CPU decode speeds of 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 while using less than 2.5GB of memory.
- The model supports a 128,000-token context window after being pretrained on approximately 34 trillion tokens and receiving an additional mid-training phase.
- Both LFM2.5-2.6B and its base version were available on Hugging Face at announcement, with support listed for llama.cpp, MLX, vLLM, SGLang, ONNX and Transformers.
Risks & concerns
- LFM2.5-2.6B scored 59.41 on LiveCodeBench v6, trailing Gemma-4-E4B-it at 63.77 and Qwen3.5-9B at 69.86, which indicates a disadvantage for coding-heavy applications.
- Its BFCLv4 tool-use score of 56.88 remained below the 60.13 achieved by the 9.7-billion-parameter Qwen3.5-9B model.
- On PinchBench, the model scored 68.22, behind Qwen3.5-4B at 71.26 and Qwen3.5-9B at 71.45.
- The claimed 30-token-per-second phone performance lacks a named device and configuration in the article, limiting the result’s reproducibility.
- All performance comparisons were published by Liquid AI, and the announcement does not include independent verification or complete details for benchmark, quantization and power-consumption settings.