Liquid AI Releases LFM2.5-2.6B and Details Its On-Device Agent Post-Training Recipe

Liquid AI released LFM2.5-2.6B on Hugging Face, a small, fast, tool-calling model optimized for agentic tasks on local hardware. Researcher helloiamleonie followed with a series of long-form posts detailing the full design of the LFM2 family (LFM2.5-350M to 2.6B) and its post-training recipe; post-training team members Maxime Labonne, Edoard Mosca and Jiahui Wang likewise emphasized that post-training is what makes on-device agentic models actually usable.

Confirmed

• Release: LFM2.5-2.6B is open for download, with the team inviting developers to run on-device agent experiments on phones and laptops. Leonie claims it matches the 3x larger Qwen3.5-9B, meaning agents can run all day for free in the background on phones, laptops, robots and cars, with multiple concurrent instances.
• Four-step post-training recipe: classic SFT (with agent trajectory data and antidoom training); SFT+RL to train multiple specialist teachers, one per skill; MOPD (multi-domain on-policy distillation) letting the student learn from all expert teachers simultaneously; and agentic RL performed directly inside the agent harness so the model becomes familiar with each harness's system prompt and built-in tools, with the overall architecture split into three parts.
• On-policy distillation (OPD): the student's own trajectories are fed to the teacher, which scores each token based on its own probabilities (measuring "surprise"). As background, Leonie recaps RLVR: extract the final answer and compare against ground truth, yielding only one sparse pass/fail signal per trajectory — inefficient for long multi-turn generation, which is where OPD's dense per-token signal helps.
• Antidoom training: detects doom loops and addresses them; per Leonie it reduced the doom-loop rate from 10% to 1%.
• On-device constraints: an iPhone's 8GB RAM is shared with the OS and apps, leaving roughly 4GB usable — one of three key considerations for on-device agent selection (from her AI Engineer Paris talk).

Unconfirmed

Claims such as "matches Qwen3.5-9B" and "doom-loop rate 10%→1%" come from the releasing team's researcher and lack independent evaluation.

Why it matters

The series turns techniques like in-harness training, multi-domain on-policy distillation and antidoom training into a reusable recipe, providing a complete reference path for building on-device agent models; if the claimed parity with a 3x larger model holds up in community reproduction, it could significantly lower the barrier to locally deployed agents.

2026-09-25 ~ 2026-09-26 · 11 related posts

• On-device agents need 3 things: ~4GB usable RAM, prefill speed, consistent quality — helloiamleonie · 2026-09-25
• [source] Liquid AI explains on-device small model design: antidoom training cuts doom loops from ~10% to 1% in LFM2.5 — helloiamleonie · 2026-09-25
• Liquid AI's "Antidoom Training" cuts model doom loops from 10% to 1% — helloiamleonie · 2026-09-25
• RLVR refresher: verifiable rewards work but the signal is sparse — helloiamleonie · 2026-09-25
• On-policy distillation explained: teachers grade every token of student trajectories — helloiamleonie · 2026-09-25
• Multi-domain on-policy distillation lets one model learn many skills at once — helloiamleonie · 2026-09-25
• Agentic RL inside the harness: how LFM2.5 trains with sandboxes and tools — helloiamleonie · 2026-09-25
• [source] Inside LFM2.5-2.6B's post-training recipe: SFT, RL, multi-domain distillation — helloiamleonie · 2026-09-25
• LFM2.5-2.6B matches Qwen3.5-9B at 3x size, unlocking on-device agents — helloiamleonie · 2026-09-25
• [source] Liquid AI ships LFM2.5-2.6B, a small fast model built for on-device agents on phones and laptops — helloiamleonie · 2026-09-25
• Liquid AI's post-training team on what makes on-device agentic models actually useful — alexcovo_eth · 2026-09-26