40% Fewer Tool Calls on Pi: LFM2.5-24B Adapters Make Agents More Cost-Efficient

Hugging Face's younes released LoRA adapters for LFM2.5-24B-A2B, targeting the two open-source harnesses Pi and Claude Code. When Liquid AI reposted it, they emphasized that LFM was designed to be customizable from the start, and that specialized intelligence is what can give more than half of enterprise AI use cases predictable economics. In the context of this release, that claim is not empty.

The adapter did not retrain the base model. It applies LoRA on LFM2.5-24B, then uses RL to tune agent behavior. LoRA adds a small set of parameters on top of the base model, while RL uses reward signals to adjust the policy. The training recipe comes from Hugging Face's multi-harness-rl space, and the dataset is SmolDataEnvs-harbor-train. In a reply, younes said the agent used Pi in a sandbox and was not trained on synthetic Pi trajectories. The entire training run was packed onto one B300, with half of the GPU running the vLLM server and the other half running training, taking about 24 hours in total.

On performance, Pi-100 steps is the best trade-off: it preserves performance relative to the base model while improving tool-call efficiency. The official claim is up to 40% fewer tool calls on Pi. The Claude Code version has not shown a comparable improvement, and the team says it will continue working on it. In the chart, the left axis is the change in pass rate, and the right axis is the change in the number of tool calls on tasks that both sides solved. The base model's pass@1 is 0.68 under Claude Code and about 0.75 under Pi; each checkpoint uses 93 to 100 held-out tasks, T=0.8, with at most 17 agent steps.

Someone asked how the 40% figure was obtained, and younes's answer was straightforward: RL, with Pi running in a sandbox, not synthetic trajectories. Another commenter said fewer tool calls is the real metric for agents: fewer calls mean lower cost and fewer chances to go off track. That is not an exaggeration. Every extra step an agent takes is another chance to spend money and take the wrong path.

On the weights side, the team released GGUF, which can plug directly into llama.cpp. younes said the Pi LoRA works smoothly on the Pi harness. Someone asked whether the adapter would slow things down in llama.cpp, and the original post did not answer. That point will have to wait for real-world testing.

Links:

• Adapter collection
• Pi GGUF weights
• Training recipe
• Training dataset
• Liquid AI original tweet
• younes original tweet

Enterprises want predictable cost per task. Large-model bills can fluctuate; specialized small models plus adapters at least narrow the variables a bit.