最新模型
Liquid AI Releases LFM2.5-24B-A2B, Expanding Long-Context and Local Agent Tooling Capabilities
On October 13, Liquid AI unveiled LFM2.5-24B-A2B and base model weights, extending context to 131,072 tokens and switching to a fixed reasoning-before-answering format. The company published local decoding and agent benchmark results, but deployers should still validate end-to-end task latency, tool parsing, and post-quantization quality.
Liquid AI released LFM2.5-24B-A2B on October 13, continuing its sparse MoE design with 24B total parameters while offering both base and post-trained weights. The A2B in the name is an approximate label: the model card lists 2.45B active parameters per token. In practice, it should not be treated as a small model that requires loading only about 2B weights. Official announcement, model card
The new version retains a 40-layer architecture, pairing short convolution blocks with GQA. Each MoE block has 64 experts, with four selected per token. The main changes are an expanded context window, from 32,768 to 131,072 tokens; a vocabulary expanded to 128,000; and an increase in pretraining data from 17T to 33T tokens. The vocabulary expansion targets non-Latin scripts, but its effect on Chinese costs and quality still needs to be measured; the improvement cannot be inferred directly from vocabulary size. Architecture and training details
There are also concrete differences for agent integration: the post-trained version always generates reasoning first, and its chat template adds <think>. By default, tool calls use a Python-style function list wrapped in special tokens; JSON can also be requested through prompting. Existing agent frameworks should therefore check how they separate reasoning blocks, parse tools, and handle stop conditions, rather than applying output handling designed for other models as-is. Tool interface
Liquid AI measured decoding speed at 184 tokens per second on an Apple M5 Max. The figure of more than 21,000 output tokens per second on a single B300 comes from a sustained-concurrency test using SGLang 0.5.22, BF16, 1,024-token inputs, and outputs of up to 256 tokens. This is aggregate serving throughput, not the speed for a single user, and it does not include tool execution time. Benchmark conditions
Agent capability comes with trade-offs: the official LiveClaw score is 48.09, above the comparison model Qwen3.6-35B-A3B at 38.19; however, its SWE-Bench Verified score is 51.00, below the other model’s 66.20. When selecting a model, teams should rerun their actual tool workflows, measure completion time including reasoning tokens, and compare success rates for quantized versions. The model card lists the lfm1.0 license; review its terms before adoption. Benchmark table, weights and license listing