Liquid AI ships LFM2.5-24B-A2B MoE with day-0 on-device support, hitting 74 tok/s on M3 Ultra
helloiamleonie · x · 2026-10-14
Liquid AI released LFM2.5-24B-A2B, an on-device MoE (24B total, 2.45B active per token, 128K context), with day-0 support in Nativ, an open-source macOS client for running local models on Apple Silicon.
Benchmarks on M3 Ultra with BF16 MLX weights:
- 4,337 tok/s prefill, 74 tok/s decode single-stream
- 205 tok/s decode with 8 concurrent requests
- 48.4 GB peak memory; still 49 tok/s decode at full 128K context
Nativ is MIT-licensed, built on MLX-VLM with Metal/unified-memory tuning, covers text/image/video/audio/code modalities, requires no accounts or subscriptions, and can expose local models to existing coding agents.
Related event: LiquidAI Releases LFM2.5-24B-A2B On-Device MoE Model(4 posts)→
More from Infra
- Ayar Labs, Wiwynn and GUC demo integrated optical compute system with live telemetry at OCP — bookwormengr · 2026-10-14
- Gavin Baker concedes some AI deals were fake, cites $1.4T commitments talking point — GavinSBaker · 2026-10-14
- Argon pushes Gemma past 120k tokens/sec on a single A100 with vLLM, a 2.9x speedup — osanseviero · 2026-10-14
- Espresso now auto-optimizes Databricks Spark jobs and configs — mattturck · 2026-10-14
- Intel Diamond Rapids motherboard shown off: 16 DDR5 slots, MRDIMM at 12800MT/s — firstadopter · 2026-10-14
- SayGM claims 70% cheaper LLM gateway with $5M ARR and crypto referral rewards — markjeffrey · 2026-10-14