LiquidAI's LFM2.5-24B-A2B Launches and Tops Benchmarks

Episode 1 · LiquidAI Releases LFM2.5-24B-A2B, an On-Device MoE Reasoning Model (2026-10-14, 13 posts)

On October 14, LiquidAI released LFM2.5-24B-A2B, the largest and final member of the LFM2.5 family—an on-device MoE reasoning model with 24B total parameters and only 2.45B activated per token. It can run multi-step agentic tasks on low-VRAM/memory devices such as laptops and PCs, targeting fast, private local agent scenarios.

Confirmed

• Specs: context extended to 131K tokens, 128K vocabulary, and 33T tokens of pretraining data; positioned as a reasoning model that is easy to fine-tune and customize (@maximelabonne).
• Open source: GGUF weights are already available on Hugging Face (@Skyline34rGt).
• Ecosystem: local inference app Nativ offers Day-0 support; per reposts, the model achieves fully local inference at 74 tok/s on an M3 Ultra.
• Multiple reposts claim performance rivaling larger MoE models (e.g., Qwen3.6-35B), though this is secondhand and no specific benchmark figures were given in the posts.

Why it matters

With only 2.45B activated parameters, inference cost and memory footprint are far below what the total parameter count would suggest, potentially enabling a 24B-class model to run multi-step agent tasks smoothly on consumer laptops. Combined with open GGUF weights and launch-day support from local inference tools like Nativ, the barrier to private on-device deployment keeps dropping.

• LiquidAI ships LFM2.5-24B-A2B: a 24B MoE with 2.45B active params for on-device agents — Skyline34rGt · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B: 2B active params, 131K context, local agent model — maximelabonne · 2026-10-14
• Liquid AI cofounder: LFM2.5-24B is what makes "AI PCs" real — JosephJacks_ · 2026-10-14
• Liquid AI releases LFM2.5-24B-A2B: 24B params, 2B active, built for local agents — JosephJacks_ · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B MoE with day-0 on-device support, hitting 74 tok/s on M3 Ultra — helloiamleonie · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B: 24B total, 2.45B active, targets Qwen3.6-35B locally — JosephJacks_ · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B: 24B total/2B active local agent model hitting 184 tps on M5 Max — JosephJacks_ · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B: 24B params with 2B active, 184 tok/s on-device agents — JosephJacks_ · 2026-10-14
• Liquid AI ships LFM2.5-24B-A2B: a 2B-active local agent model hitting 21K tok/s on one B300 — JosephJacks_ · 2026-10-14
• Liquid AI releases LFM2.5-24B-A2B: 2B-active local reasoning model for on-device agents — HankYeomans · 2026-10-15
• Liquid AI releases LFM2.5-24B-A2B: 24B params, 2B active, 131K context, open weights — JosephJacks_ · 2026-10-15
• Liquid AI releases LFM2.5-24B: just 2B active params per token for local agents — helloiamleonie · 2026-10-15
• Liquid AI ships LFM2.5-24B-A2B: 24B params with 2B active, runs on a 16GB GPU — JosephJacks_ · 2026-10-15

Episode 2 · Liquid AI's LFM2.5 Tops Agon and IFEval Leaderboards (2026-10-15, 2 posts)

Liquid AI's LFM2.5-24B-A2B, with only 2.45B active parameters, topped the Agon benchmark with 93.4%, dethroning Qwen3.6-35B-A3B, and also reached No.1 on the IFEval leaderboard with 96.75.

• Liquid AI's LFM2.5-24B, 30-of-40 layers convolutional, dethrones Qwen 3.6 atop Agon bench at 93.4% — JosephJacks_ · 2026-10-15
• Liquid AI's LFM2.5-24B-A2B tops IFEval with just 2.45B active parameters — huggingface · 2026-10-15