Liquid AI ships LFM3: on-device 1.7B model beats Qwen3.5 9B on agent benchmarks
haider1 · x · 2026-10-28
Liquid AI released LFM3, a new generation of Liquid Foundation Models with a tri-hybrid architecture and multimodal encoders: six open-weight models around 2B parameters spanning text, vision, audio, and decision-making.
- LFM3-1.7B: text, 256k context window
- LFM3-VL-2B: vision-language with video input for spatial understanding
- LFM3-Audio-2B: tool calls, voice chat, reasoning in 18 languages
- LFM3-Omni-2B: combined vision, audio, and language
- d1-2B / d1-Omni-2B: decision models with typed outputs and calibrated probabilities in a single pass
Benchmarks show the 1.7B model beating Qwen3.5-9B on WebArena-Verified (38.4), LiveClaw (45.6), and AutomationBench, plus leading tool-use scores on BFCL V4 and ToolSandbox; it trails on SWE-bench Verified (43.4 vs 51.2).
On-device (llama.cpp Q40, 32K context), LFM3-1.7B was the fastest model on AMD Ryzen AI Max 395, Apple M5 Max, and iPhone 17 Pro, decoding 1.8-2.8x faster than the next model with the lowest memory use; a full 128K context needs only 1.6 GB of KV cache vs 4.3 GB for Qwen3.5-4B. Note all numbers are Liquid AI's own tests.
More from Infra
- Cotidal emerges to build AGI supercomputers, backed by TPU/LLVM/HF luminaries — Ar_Douillard · 2026-10-28
- Petabyte pitches sleep-when-idle hosting for 24/7 AI agents with event-triggered wakeups — Petabyte-Cloud · 2026-10-28
- Run a coding agent in your browser: LFM3-1.7B hits 300+ tok/s on M4 Max via WebGPU — victormustar · 2026-10-28
- Runpod's BenchPod: Stop Guessing vLLM Configs, Cut Kimi K3 Latency on B300 — AI Engineer · 2026-10-28
- Ex-Intel CEO Pat Gelsinger bets on Nagare Labs' Z2RAM to vault the memory wall — zephyr_z9 · 2026-10-28
- Nvidia pushing NVHBM to de-risk base die after Hynix and Micron fumble, says report — basedjensen · 2026-10-28