Liquid AI Unveils MoE Reasoning Model 'LFM2.5-24B-A2B', Reaching 184 Tokens/s on Apple M5 Max

On October 13, 2026, Liquid AI officially introduced 'LFM2.5-24B-A2B', a new Mixture of Experts (MoE) reasoning model that activates only 2.45 billion (2.45B) parameters per generated token out of its total 24 billion (24B) parameter footprint, making model weights publicly available on Hugging Face. Engineered as a dedicated reasoning model that executes an inherent structured thinking phase before formulating final responses, the release includes MLX quantized weights tailored for Apple Silicon and published local benchmark figures.

Image source: Liquid AI / X (@koreabuldak)

The most prominent technical metric highlighted in the release centers on on-device and desktop workstation inference throughput. According to Liquid AI's Hugging Face model card, the model recorded an inference speed of 184 tokens per second (tokens/s) when running 4-bit quantization under llama.cpp on Apple's M5 Max chip.

MoE Reasoning Architecture: 2.45B Active Out of 24B Parameters

The 'A2B' designation in the LFM2.5-24B-A2B model identifier signifies that the active parameter count required per token generation is constrained to approximately 2.45 billion.

While the aggregate parameter pool reaches 24 billion, the architecture employs a sparse Mixture of Experts (MoE) routing mechanism, selectively invoking only the relevant expert sub-networks based on the context of each incoming token.

• Compute and Memory Bandwidth Efficiency: Standard dense 24B models must stream all 24 billion weights from memory for every single generated token. In contrast, LFM2.5-24B-A2B accesses only the 2.45B active parameters per token, drastically cutting required memory bandwidth and unlocking exceptionally high generation speeds on unified memory architectures.
• Dedicated Thinking Phase: Unlike general autoregressive base models that produce direct immediate outputs, LFM2.5-24B-A2B is trained specifically as a reasoning model that executes an explicit thinking workflow before emitting final answers, optimizing it for structured multi-step logic and complex problem-solving.

Apple M5 Max Benchmark: 184 Tokens/s and MLX Memory Requirements

While the 184 tokens/s throughput figure demonstrates the potential of sparse MoE inference on Apple Silicon, the metric reflects specific testing parameters that must be evaluated alongside hardware constraints.

• Benchmark Configuration: The 184 tokens/s result was benchmarked on an Apple M5 Max processor running llama.cpp with 4-bit quantization (Q4_K_M) across a 4K input context window.
• MLX Quantization Weight Sizes: In the MLX ecosystem optimized for macOS, the model is distributed in a 4-bit variant requiring 13.6GB of storage and an 8-bit variant requiring 25.5GB.
• Unified Memory Footprint Constraints: For Mac systems equipped with 16GB of unified memory (such as entry-level Mac mini setups), loading the 13.6GB 4-bit model leaves an extremely slim operating buffer once macOS and base system services claim their share of RAM. Deploying the higher-precision 8-bit model (25.5GB) necessitates hardware with at least 32GB of unified memory.
• Precision vs. Bandwidth Trade-off: Running the 8-bit version roughly doubles the active weight data volume read across memory channels during token generation. As a consequence, memory bandwidth saturation prevents the 8-bit build from sustaining the 184 tokens/s observed on the Q4_K_M benchmark, resulting in proportionally lower generation speeds.

Official Korean Language Support and RAG Guidelines for Knowledge-Intensive Tasks

LFM2.5-24B-A2B also includes Korean in its roster of officially supported languages, allowing developers and researchers to deploy local reasoning workflows without requiring auxiliary translation layers or custom fine-tuning.

• Multilingual Support: With Korean explicitly supported out of the box, developers can integrate native Korean prompting and system directives into local pipelines.
• Retrieval-Augmented Generation (RAG) Recommendation: The Hugging Face model card explicitly advises that for knowledge-intensive domains or tasks demanding precise factual lookup, the model should be paired with external search tools or Retrieval-Augmented Generation (RAG) pipelines rather than relying solely on parametric memory. This hybrid setup pairs the model's structural reasoning capabilities with grounded external real-time data.

Sources

• Hugging Face Model Repository: LiquidAI/LFM2.5-24B-A2B
• 갑오징어 X (@koreabuldak): 2026-10-14 LFM2.5-24B-A2B Benchmark and Local Memory Analysis