Liquid AI's new open-weight model runs agentic reasoning locally with just 2B active parameters
In short: Liquid AI has published LFM2.5-24B-A2B, a mixture-of-experts model with 24B total parameters but only 2B active per token, designed to run reasoning agents on local hardware. The release expands the context window to 131K tokens, doubles the vocabulary size, and was trained on roughly double the pretraining data of its predecessor. On agentic and tool-use benchmarks the model performs on par with or better than models that have about 50% more parameters, while delivering fast inference speeds on both consumer and data-center hardware.
This summary was generated automatically by AI from Hugging Face's publication. It is our own text, not a copy of the original — facts, figures and quotes belong to the source, linked above and below.
What changed?
- 1Context window increased from 32K to 131K tokens
- 2Vocabulary size doubled from 65K to 128K tokens, improving tokenization for non-Latin scripts
- 3Pretraining data scaled from roughly 17T to 33T tokens, with added mid-training on reasoning, math, code, and multilingual text
- 4Model now produces explicit chain-of-thought reasoning before answering, unlike its predecessor
- 5Post-training uses five stages including supervised fine-tuning, teacher distillation, and reinforcement learning inside live agent harnesses
- 6Supported out of the box in llama.cpp, MLX, vLLM, SGLang, and ONNX
- 7Both base and instruct-tuned weights released openly on Hugging Face
| Parameter | Before | Now |
|---|---|---|
| Context window | 32K tokens | 131K tokens |
| Vocabulary size | 65K tokens | 128K tokens |
| Pre-training tokens | ~17T tokens | ~33T tokens |
| Reasoning | Not reasoning-only | Reasoning-only with explicit chain of thought |
| Active parameters | 2B (predecessor LFM2-24B-A2B) | 2B (unchanged) |
| Tool calling | Not specified | Supported, trained inside agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode) |
| Inference speed (Apple M5 Max) | Not specified | 184 tok/s |
| Inference speed (NVIDIA B300) | Not specified | Up to 21K output tok/s |
Why it matters
For teams building tool-calling agents, a model that runs on local hardware with low active-parameter cost but still competes with much larger models on agentic benchmarks (like LiveClaw and AutomationBench) means cheaper, private, always-on automation without relying on a cloud API for every reasoning step.
What it means for AI agents and contact centers
For voice AI and contact-center automation, the model's strong tool-use and instruction-following scores combined with low inference cost make it worth testing as a local backend for agent harnesses that need to call functions, follow multi-step workflows, or operate without sending data to external APIs; its multilingual tokenizer upgrade is also relevant for non-English language handling.
🧪 Worth Testing
The combination of strong tool-use benchmark results, a 131K context window, and fast local inference (184 tok/s on consumer-grade Apple silicon) makes this a practical candidate to benchmark against cloud-hosted agent models for cost, latency, and privacy in automation pipelines.
Sources
- Hugging FaceOfficialPrimary sourceOriginal article →„Run fast, private agents on-device with LFM2.5-24B-A2B“13 Oct 2026, 20:05
- Published by source
- 13 Oct 2026, 20:05
- Found by our system
- 13 Oct 2026, 20:11
- Summary generated
- 13 Oct 2026, 20:13
This article was written by AI from the original source. Facts, numbers and prices come from the source; missing values are marked “Not specified”. Legal notice, copyright and privacy