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

Liquid AI has released the on-device model LFM2.5-2.6B, now available for download on Hugging Face. It is positioned as a compact, fast, locally deployable model optimized for agentic tasks. Liquid AI researcher helloiamleonie then published a series of long-form posts systematically explaining the complete design methodology and post-training recipe for the LFM2 family of small on-device models (LFM2.5-350M to 2.6B)—dense with information and a first-hand resource for understanding agent training on small models.

Confirmed

• Model release: LFM2.5-2.6B is open for download; the team is calling on developers to run it on phones or laptops for on-device Agent experiments, with tool-calling support. Leonie claims it matches Qwen3.5-9B at 3x its size, and stresses that agents can thus run locally, for free, around the clock in the background on phones, laptops, robots, and cars, with multiple concurrent instances.
• A four-step post-training recipe: classic SFT (with agent trajectory data and antidoom training); SFT+RL to train multiple specialized teachers for individual skills; MOPD (multi-domain on-policy distillation) so the student learns from domain-expert teachers simultaneously; and an agentic RL stage trained directly inside the agent harness, exposing the model early to each harness's system prompt and built-in tools.
• On-policy distillation (OPD) principle: feed the student's own trajectories to the teacher, which scores token-by-token based on its own probabilities (measuring 'surprise'); compared to RLVR's sparse reward of a single pass/fail per trajectory, the signal is much denser.
• Antidoom training: detects doom loops and addresses them specifically, cutting the loop rate from 10% to 1% per Leonie.
• On-device constraints: an iPhone's 8GB RAM, shared with system apps, leaves roughly 4GB usable—one of three key factors in on-device Agent selection (this content is a threaded version of a talk given at AI Engineer Paris).

Why it matters

• The series turns engineering practices like 'training inside the harness,' 'multi-domain on-policy distillation,' and 'antidoom training' into a reusable, public recipe, providing a complete reference path for on-device small models doing agent work; if the claim of matching models 3x larger holds up in community replication, the barrier to locally deployed agents drops significantly.

Not yet confirmed

• Claims like 'matching Qwen3.5-9B' and 'doom loop rate 10%→1%' all come from the releasing company's own researcher and lack independent evaluation.

2026-09-25 ~ 2026-09-25 · 10 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
• Inside LFM2.5-2.6B's post-training recipe: SFT, RL, multi-domain distillation — helloiamleonie · 2026-09-25
• [source] 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