{"id":43624,"date":"2025-10-11T06:57:42","date_gmt":"2025-10-11T06:57:42","guid":{"rendered":"https:\/\/youzum.net\/liquid-ai-releases-lfm2-8b-a1b-an-on-device-mixture-of-experts-with-8-3b-params-and-a-1-5b-active-params-per-token\/"},"modified":"2025-10-11T06:57:42","modified_gmt":"2025-10-11T06:57:42","slug":"liquid-ai-releases-lfm2-8b-a1b-an-on-device-mixture-of-experts-with-8-3b-params-and-a-1-5b-active-params-per-token","status":"publish","type":"post","link":"https:\/\/youzum.net\/ja\/liquid-ai-releases-lfm2-8b-a1b-an-on-device-mixture-of-experts-with-8-3b-params-and-a-1-5b-active-params-per-token\/","title":{"rendered":"Liquid AI Releases LFM2-8B-A1B: An On-Device Mixture-of-Experts with 8.3B Params and a 1.5B Active Params per Token"},"content":{"rendered":"

How much capability can a sparse 8.3B-parameter<\/strong> MoE with a ~1.5B active path<\/strong> deliver on your phone without blowing latency or memory? Liquid AI has released <\/strong>LFM2-8B-A1B<\/strong>,<\/a> a small-scale Mixture-of-Experts (MoE) model built for on-device execution under tight memory, latency, and energy budgets. Unlike most MoE work optimized for cloud batch serving, LFM2-8B-A1B targets phones, laptops, and embedded systems. It showcases 8.3B total parameters<\/strong> but activates only ~1.5B parameters per token<\/strong>, using sparse expert routing to preserve a small compute path while increasing representational capacity. The model is released under the LFM Open License v1.0 (lfm1.0)<\/strong><\/p>\n

Understanding the Architecture<\/strong><\/h3>\n

LFM2-8B-A1B retains the LFM2 \u2018fast backbone\u2019 and inserts sparse-MoE feed-forward blocks to lift capacity without materially increasing the active compute. The backbone uses 18 gated short-convolution blocks<\/strong> and 6 grouped-query attention (GQA) blocks<\/strong>. All layers except the first two<\/strong> include an MoE block; the first two remain dense for stability. Each MoE block defines 32 experts<\/strong>; the router selects top-4 experts per token<\/strong> with a normalized-sigmoid gate<\/strong> and adaptive routing bias<\/strong> to balance load and stabilize training. Context length is 32,768 tokens<\/strong>; vocabulary size 65,536<\/strong>; reported pre-training budget ~12T tokens<\/strong>. <\/p>\n

This approach keeps per-token FLOPs and cache growth bounded by the active path (attention + four expert MLPs), while total capacity allows specialization across domains such as multilingual knowledge, math, and code\u2014use cases that often regress on very small dense models.<\/p>\n

\n
\"\"
https:\/\/www.liquid.ai\/blog\/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts<\/figcaption><\/figure>\n<\/div>\n

Performance signals<\/strong><\/h3>\n

Liquid AI reports that LFM2-8B-A1B runs significantly faster than Qwen3-1.7B<\/strong> under CPU tests using an internal XNNPACK-based stack and a custom CPU MoE kernel. The public plots cover int4 quantization with int8 dynamic activations<\/strong> on AMD Ryzen AI 9 HX370<\/strong> and Samsung Galaxy S24 Ultra<\/strong>. The Liquid AI team positions quality as comparable to 3\u20134B dense models<\/strong>, while keeping the active compute near 1.5B<\/strong>. No cross-vendor \u201c\u00d7-faster\u201d headline multipliers are published; the claims are framed as per-device comparisons versus similarly active models.<\/p>\n

On accuracy, the model card lists results across 16 benchmarks, including MMLU\/MMLU-Pro\/GPQA (knowledge), IFEval\/IFBench\/Multi-IF (instruction following), GSM8K\/GSMPlus\/MATH500\/MATH-Lvl-5 (math), and MGSM\/MMMLU (multilingual). The numbers indicate competitive instruction-following and math performance within the small-model band, and improved knowledge capacity relative to LFM2-2.6B, consistent with the larger total parameter budget.<\/p>\n

\n
\"\"
https:\/\/www.liquid.ai\/blog\/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts<\/figcaption><\/figure>\n<\/div>\n
\n
\"\"
https:\/\/www.liquid.ai\/blog\/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts<\/figcaption><\/figure>\n<\/div>\n

Deployment and tooling<\/strong><\/h3>\n

LFM2-8B-A1B ships with Transformers\/vLLM for GPU inference and GGUF builds for llama.cpp; the official GGUF repo lists common quants from Q4_0 \u22484.7 GB<\/strong> up to F16 \u224816.7 GB<\/strong> for local runs, while llama.cpp<\/strong> requires a recent build with lfm2moe<\/code> support (b6709+<\/strong>) to avoid \u201cunknown model architecture\u201d errors. Liquid\u2019s CPU validation uses Q4_0<\/strong> with int8 dynamic activations<\/strong> on AMD Ryzen AI 9 HX370<\/strong> and Samsung Galaxy S24 Ultra<\/strong>, where LFM2-8B-A1B shows higher decode throughput than Qwen3-1.7B<\/strong> at a similar active-parameter class; ExecuTorch<\/strong> is referenced for mobile\/embedded CPU deployment.<\/p>\n

\n
\"\"
https:\/\/www.liquid.ai\/blog\/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts<\/figcaption><\/figure>\n<\/div>\n
\n
\"\"
https:\/\/www.liquid.ai\/blog\/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts<\/figcaption><\/figure>\n<\/div>\n

Key Takeaways<\/strong><\/h3>\n