2026-10-21
LFM2.5-24B-A2B, a 24B agent model with 2.45B active
Liquid AI publishes LFM2.5-24B-A2B on Hugging Face. It is a reasoning model with 24 billion parameters, 2.45 billion active, a 131,072-token context, and the LFM Open License v1.0.
On 13 October 2026 Liquid AI published LFM2.5-24B-A2B. The weights are on Hugging Face, in LiquidAI/LFM2.5-24B-A2B. As of 21 October 2026 that repository has 531 likes and 16,939 downloads. The official GGUF, LiquidAI/LFM2.5-24B-A2B-GGUF, is the one people download: 227 likes and 143,838 downloads.
What it is
It is the last model in the LFM2.5 family, which starts at LFM2.5-230M. The model card lists 24.0 billion total parameters and 2.45 billion active. It is text only, and it is a reasoning model: it writes a chain of thought before the answer. The <think> tag is already in the chat template.
Two things change from LFM2-24B-A2B. Context grows from 32,768 to 131,072 tokens. The vocabulary doubles, from 65,000 to 128,000 tokens, for non-Latin languages. The listed languages are English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish.
| Layers | 40, of which 30 short convolutions and 10 GQA attention blocks |
| Experts | 64 per MoE block, 4 active per token |
| Context | 131,072 tokens |
| Vocabulary | 128,000 tokens |
| Training budget | 36 trillion tokens, as written on the card |
Liquid's post splits training into about 33 trillion pretraining tokens, then 2 trillion at 32K context and 300 billion at 131K. Post-training has five stages and includes reinforcement learning inside agent harnesses, among them OpenClaw, Hermes Agent, Pi, and OpenCode.
The sampling settings recommended on the card are temperature 0.1, top_k 50, and repetition penalty 1.05.
The numbers Liquid published
Liquid evaluated the model with vLLM. It leads or matches on five of eight tool and agent benchmarks. Two wide gaps, and the coding benchmark where the larger model stays ahead, all against Qwen3.6-35B-A3B:
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B |
|---|---|---|
| AutomationBench | 27.69 | 8.72 |
| LiveClaw | 48.09 | 38.19 |
| SWE-Bench Verified, with OpenCode | 51.00 | 66.20 |
For knowledge-heavy work, Liquid recommends pairing the model with retrieval or web search.
The speed figures are Liquid's, collected with Pipette. On-device decode is 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395. On a single NVIDIA B300, in BF16, with SGLang 0.5.22, 1,024-token prompts and up to 256 output tokens, aggregate throughput passes 21,000 output tokens per second at high concurrency. These are not SophyAI measurements, and the B300 figure is not a single-user tokens/s.
License
The license is LFM Open License v1.0, marked other on Hugging Face. Commercial use is granted to entities under an annual-revenue threshold of 10 million US dollars. Above that threshold the license does not grant commercial use.
In the catalog
Two reference profiles, both with launch=reference and launchable=false. The checkpoint is the official BF16, about 48 GB.
| Profile | GPU | Catalog band |
|---|---|---|
lfm25-24b-vllm-b300 | 1× B300 | 10–30 USD/h |
lfm25-24b-vllm-pro6000 | 1× RTX PRO 6000 | 1.2–2.0 USD/h |
The first matches the machine in Liquid's GPU benchmark. The second is there because the BF16 weights fit in 96 GB of VRAM. Liquid does not publish a tokens/s figure for the PRO 6000. The server on these profiles is vLLM. The 21,000 tokens/s figure is from SGLang, on the same GPU class.
The Hub also has the base LFM2.5-24B-A2B-Base, the FP8, GGUF files for llama.cpp, MLX checkpoints for Apple, and ONNX. The official Q4_K_M weighs 14.5 GB.