Liquid AI ships LFM2.5-24B-A2B: 24B params with 2B active, runs on a 16GB GPU
JosephJacks_ · x · 2026-10-15
Liquid AI released LFM2.5-24B-A2B, the largest LFM2.5 model: 24B total / 2B active params per token, 131K context, 128K vocab, 33T pretraining tokens, a reasoning model aimed at local multi-step agentic tasks on laptops/PCs, with open base and post-trained weights. It claims competitiveness with Qwen3.6-35B-A3B (50% more params) on agentic/tool-use tasks.
Community benchmarks on the same Q4 GGUF (14.5GB, fits 16GB cards) via llama.cpp: RTX 5090 vs 16GB RTX 5060 Ti — math 1.98s vs 3.99s (both 450/450 correct), writing 351.5 vs 162.0 tok/s, GSM8K 97% on both, MMLU 80% vs 82%. Catch: thinking is always on.
Related event: LiquidAI Releases LFM2.5-24B-A2B, an On-Device MoE Reasoning Model(13 posts)→
More from Models
- Nathan Lambert slams Meta's Muse Open License as 'trash' — natolambert · 2026-10-16
- Google launches Gemini 4 Argon to Ultra and Pro subscribers, touting physics and 3D skills — ShikharMurty · 2026-10-16
- Google gives X users early access to test Argon, not just companies — lyraxana · 2026-10-16
- NVIDIA and PrimeIntellect post-train Nemotron 3.5 Super VL, boosting AutomationBench-AA by ~70% — willcb · 2026-10-16
- Grok Voice 3.0 Tops Design Arena Speech Leaderboard, Up 7.4 Points — XFreeze · 2026-10-16
- Robotics LLMs tested on 16 tasks: GPT-6 Astra hits 17.2% success vs Fable's 9.1% — pabbeel · 2026-10-16