Liquid AI's New LFM 2.5 Is INSANE — 184 TOK/S With Only 2B Active! (Live Tests)

Liquid AI LFM 2.5 A2B: Hybrid Engine for Complex Agentic Tasks

This video reviews Liquid AI’s LFM 2.5 A2B, a hybrid engine that activates only 2B of 24B total parameters per task. The presenter covers architecture, benchmarks, local deployment, and demonstrates six practical use cases on playground.liquid.ai.

Architecture & Performance

• Mixture-of-experts routing selects 4 of 64 experts per task; double-gated convolutional layers replace most transformer blocks, cutting 75% of layers.
• Supports a 100K-token context window, 128K vocabulary, and 16 languages; trained on 36T tokens.
• Temperature is locked at 0.1–0.2 for deterministic output; a direct chain-of-thought process keeps tool execution precise. Speeds reach 184 tok/s on Apple M5 and 64 tok/s on CPU-assisted setups (1.7× faster).

Benchmark Highlights

• 48 on agentic harness, beating Qwen 35B, Google Gemma, and Nvidia Nemotron by 26%; 28% lab automation (3.2× higher), 97% instruction compliance, >80% function calling.
• Lags slightly in scientific reasoning, mathematical proofs, and code repair versus Qwen 35B.

Hands-On Test Results

• Inventory analysis: produced a table, bar chart, and a 10×85 heat-map grid with unique symbols.
• Project dashboard: single-file HTML tracker with sidebar, dynamic table, and clean CSS.
• Conway’s Game of Life: terminal Python glider/blinker with correct boundary wrapping.
• Kanban board: single-file HTML app with add and “move right” status transitions.
• Speedometer: native vector diagram with red/yellow/green zones, needle at 78%; arc shape imperfect.
• Sine wave animation: terminal star character output with two full cycles.

Local Deployment

• Free online at playground.liquid.ai; locally via llama.cpp and GGML files.
• 4-bit quantized: 14 GB, 12–16 GB VRAM; offload version: 6 GB VRAM + 16 GB RAM.
• 2-bit mobile: ~10 GB download, 8–12 GB VRAM.
• Set repeat penalty to 1.05; enable flash attention for large contexts.

Pros and Caveats

• ✅ High speed, low dependency overhead, transparent benchmarks.
• ⚠️ Temperature above 0.5 causes hallucinations; not for creative prose; uses Liquid AI’s proprietary open-weight license, not MIT.

Key Takeaway

“We do not need massive computing power to get best results on agentic tasks.”

LFM 2.5 A2B brings powerful agentic reasoning to budget hardware, but only within strict low-temperature, analytical settings.