LFM2.5 2.6B vs Qwen 3.5: local tests and llama.cpp setup
Learn how LFM2.5 2.6B runs through llama.cpp and compares with Qwen 3.5 4B in reasoning, tool calling and an OpenCode newsletter app task.
Player not loading? Watch on YouTube
This review tests Liquid AI's LFM2.5 2.6B dense model as a local LLM, with a short llama.cpp server setup and an OpenCode coding task. The presenter describes base and instruction-tuned releases, a 128K context window and GGUF downloads in several precisions. He uses the 8-bit version with recommended sampling settings. Phone and CPU responsiveness are stated goals rather than results established by these tests.
The presenter reports roughly 3.5 GB of RAM use and 80 tokens per second during recording, rising to about 100 when he is not recording. He identifies his machine as an M5 Pro with 48 GB of unified memory. Results vary: the model incorrectly recommends walking to a car wash without the car, but correctly answers a falling-objects question set in a vacuum.
In his benchmark suite, Qwen 3.5 4B at 4-bit quantization outperforms LFM2.5 at 8-bit, especially in tool calling. He reports similar generation speeds, with Qwen using about 10 to 12 GB more memory. These findings describe his configuration and tests.
The coding assistant trial asks OpenCode to build a small newsletter platform. Invalid tool calls and repeated path errors precede a rendered interface whose subscriber action does nothing in the demonstration. The presenter questions how published benchmarks translate to his tasks, while acknowledging that his evaluation may differ from the model's training focus.