LFM2.5-24B Adapter: From Launch to Local GGUF

Episode 1 · LFM2.5-24B Harness Adapters Cut Tool Calls ~39% via LoRA+RL (2026-10-30, 6 posts)

The Liquid AI team (with training pipeline details from QGallouedec) has released a "LFM2.5-24B Harness adapters" collection on Hugging Face's harness-hub, offering LoRA adapters for LFM2.5-24B-A2B tailored to two open-source agent frameworks—Pi and Claude Code—along with GGUF versions. Training uses LoRA + RL: first bootstrapping from a strong model, then reinforcement learning, all doable on a single GPU in 24 hours.

Confirmed

• The collection includes both Pi and Claude Code LoRA adapters plus GGUF versions, now live on Hugging Face
• Official claim: 39–40% fewer tool calls in the pi harness with no loss in accuracy/success rate
• An independent comparison by @yb2698: on pass@1, the Pi run (lr 1e-5) dropped 8 percentage points below baseline at 110 steps, with comparison data for the claude-code run also posted; overall takeaway is that 100 steps is the sweet spot for Pi, yielding roughly a 37% reduction in tool calls
• Training cost: 24 hours on a single GPU

Why it matters

• The work shows that low-cost RL fine-tuning can meaningfully improve the runtime efficiency of open-source agent frameworks by cutting unnecessary tool calls, directly lowering inference costs for agentic development
• With adapters customized per harness and GGUF versions included, deployment on local setups and community replication are easy, offering the open-source agent ecosystem a reusable template for efficiency optimization

• harness-hub ships LFM2.5-24B LoRA adapter collection for Pi and Claude Code — yb2698 · 2026-10-30
• LFM2.5-24B LoRA adapters for Pi and Claude Code cut tool calls up to 40% — yb2698 · 2026-10-30
• Pi-100 steps hit best tradeoff: LFM2.5 LoRA cuts 37% tool calls — yb2698 · 2026-10-30
• LFM2.5-24B LoRA adapters for Pi and Claude Code cut tool calls by up to 40% — yb2698 · 2026-10-30
• RL-Trained LoRA Cuts LFM2.5-24B Tool Calls by 39% in Agent Harnesses — amaarora · 2026-10-30
• LFM2.5-24B LoRA adapters for Pi and Claude Code cut tool calls by up to 40% — yb2698 · 2026-10-30

Episode 2 · LFM2.5 Pi LoRA Adapter Released in GGUF for Local llama.cpp Use (2026-10-30, 2 posts)

The team released a GGUF version of the LFM2.5-24B-A2B Pi LoRA agent adapter (only 37MB in F16), enabling local execution via llama.cpp. Tests on the Pi harness show good results, with more harnesses planned.

• LFM2.5 Pi LoRA goes GGUF: run the RL-trained agent adapter in llama.cpp — yb2698 · 2026-10-30
• GGUF weights released, plug into llama.cpp with Pi LoRA on Pi harness — yb2698 · 2026-10-30