Liquid AI API

LFM2 24B A2B is a sparse Mixture-of-Experts model from Liquid AI featuring a novel hybrid architecture that combines gated short convolution blocks with Grouped Query Attention in a 3:1 ratio, developed through hardware-in-the-loop architecture search. With 24 billion total parameters but only ~2 billion active per token, it delivers high throughput while outperforming larger MoE competitors like Qwen3-30B-A3B in throughput benchmarks. It supports 9 languages, a 32K context window, native function calling, and structured outputs. A strong API choice for high-volume multi-agent pipelines, RAG backends, and multilingual applications that demand low per-token cost alongside capable general reasoning.

LFM 2.5 1.2B Instruct is a compact instruction-tuned language model from Liquid AI, designed to deliver best-in-class performance at the 1-billion-parameter scale. Trained on 28 trillion tokens with reinforcement learning, it achieves strong scores across knowledge (MMLU-Pro: 44.35), reasoning (GPQA: 38.89), and instruction following (IFEval: 86.23) — outperforming similarly sized models like Llama-3.2-1B and Gemma-3-1B on these benchmarks. The model supports tool use, structured outputs, and function calling, making it a solid choice for lightweight agentic pipelines, chatbots, and latency-sensitive API integrations where cost and throughput matter most.

LFM 2.5 1.2B Thinking is a compact reasoning model from Liquid AI that generates explicit chain-of-thought traces before producing answers, enabling more reliable performance on multi-step problems at the 1-billion-parameter scale. Compared to its instruct sibling, it shows major benchmark gains in math reasoning (MATH-500: 88 vs. 63), instruction following (Multi-IF: 69 vs. 61), and tool use (BFCLv3: 57 vs. 49). It matches or exceeds Qwen3-1.7B on most reasoning benchmarks despite having 40% fewer parameters. Well-suited for API use cases involving agentic tool calling, math, and code — anywhere a reasoning trace meaningfully improves answer quality.

LFM2-8B-A1B is a sparse Mixture-of-Experts language model from Liquid AI with 8.3B total parameters but only 1.5B active per token, using 32 experts per MoE block with top-4 active per token. This design delivers 3-4B dense model quality at the compute cost of a 1.5B model, making it faster than Qwen3-1.7B in practice. Verified benchmarks include GSM8K 84.4%, MATH500 74.2%, IFEval 77.6%, and MMLU-Pro 37.4%. For API developers, it is a strong choice for latency-sensitive applications requiring larger-model quality at minimal compute cost — ideal for high-throughput pipelines where speed and efficiency are priorities.

LFM2-2.6B is a hybrid language model from Liquid AI, built on a novel architecture that alternates Grouped Query Attention blocks with gated short convolutional layers. Trained on 10 trillion tokens, it delivers fast inference with a significantly reduced KV cache footprint compared to pure-transformer models. Despite its 2.6B parameter count, it outperforms larger models in its class including Llama 3.2-3B-Instruct and Gemma-3-4b-it. Verified benchmarks include 82.41% on GSM8K and 79.56% on IFEval, surpassing Llama 3.2-3B's 71.43% on the latter. For API developers, it is well-suited for low-latency, cost-efficient inference tasks such as instruction following, Q&A, and math-related applications.