Liquid AI's LFM2.5-24B, 30-of-40 layers convolutional, dethrones Qwen 3.6 atop Agon bench at 93.4%
JosephJacks_ · x · 2026-10-15
Liquid AI's LFM2.5-24B-A2B took the top spot on the Agon benchmark with 185/198 (93.4%), dethroning Qwen3.6-35B-A3B (89.9%). Key facts:
- Own architecture, not a Qwen finetune: 30 of 40 layers run as convolutions instead of attention; 24B total / 2B active per token, 64 experts routed top-4, and a 128K tokenizer expanded in place from LFM2's 65K
- Sub-scores: Agent 91.7%, code 93.3%, reasoning 96.1% — first of 23 models with complete runs
- Local-friendly: 14.5 GB at Q4KM, fits an 8GB card with expert layers offloaded to CPU
- The author notes a config pitfall: stock card defaults caused some failures until offload settings were adjusted
Related event: Liquid AI's LFM2.5 Tops Agon and IFEval Leaderboards(2 posts)→
More from Infra
- engy.ai processes 50B+ tokens in a day as open-model inference demand surges — const_reborn · 2026-10-16
- New inference engine 'e' hits 10,500 tok/s on a single DGX Station — brandon_galang · 2026-10-16
- SGLang ships Day-0 support for FLUX 3 Image Klein: 10 reference images, BF16/FP8/NVFP4 variants — ying11231 · 2026-10-16
- NVIDIA's Nemotron 3.5 Super VL hits CoreWeave: 12B-active multimodal MoE at $0.25/M input tokens — wandb · 2026-10-16
- Chip equipment beats bull case: 2026 WFE seen at $153B, $278B by 2030 — BenBajarin · 2026-10-16
- Steam now launches in 5.23s, beating Epic's 5.47s — and vLLM gets roasted for slow startup — AlpinDale · 2026-10-16