Liquid AI · 2026-10-13 · major
LFM2.5-24B-A2B — Liquid AI's biggest on-device agent model
Liquid AI released LFM2.5-24B-A2B, an open-weight 24B MoE with about 2B active parameters and a 128K context. It scores 27.69 on AutomationBench, over three times Qwen3.6-35B-A3B, and runs at 184 tok/s on an Apple M5 Max.
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png)
A 24B open-weight MoE that wakes only ~2B parameters per token, built to run agents on laptops.
Key specs
| Automation bench | 27.69 |
|---|---|
| Speed (m5 max) | 184 tok/s |
Quick facts
| Maker | Liquid AI |
|---|---|
| Parameters | 24B total, ~2B active (64 experts, top-4) |
| Context window | 131,072 tokens (up from 32K) |
| Pre-training | 33T tokens (up from 17T) |
| Mode | Reasoning-only (thinks before answering) |
| Availability | Open weights; GGUF, MLX, vLLM, SGLang, ONNX |
| API price | $0.10 in / $1.00 out per 1M tokens |
Benchmarks
| LFM2.5-24B-A2B | 27.69 | |
|---|---|---|
| Qwen3.6-35B-A3B | 8.72 | |
| Gemma-4-26B-A4B-it | 5.46 | |
| Nemotron-3.5-Lightning 30B-A3B | 2.61 |
| LFM2.5-24B-A2B | 51% | |
|---|---|---|
| Qwen3.6-35B-A3B | 66.2% | |
| Gemma-4-26B-A4B-it | 25.8% | |
| Nemotron-3.5-Lightning 30B-A3B | 53.4% |
Pricing
| Input | $0.10 / 1M tokens |
|---|---|
| Cached input | $0.05 / 1M tokens |
| Output | $1.00 / 1M tokens |
What is it?
LFM2.5-24B-A2B is the largest and final model in Liquid AI's LFM2.5 family: a text-only, reasoning-only model for private agents that run on your own device. Liquid AI ships base and post-trained weights on Hugging Face, plus first-party GGUF and MLX builds.
How does it work?
The mixture-of-experts design has 40 layers that mix gated short convolutions with grouped-query attention, and each block routes a token to 4 of 64 experts, so only about 2B of 24B parameters do work per token. Post-training runs five stages, from SFT and teacher distillation to agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
Why does it matter?
Agent work like tool calls and multi-step automation can now run locally at high speed, with data staying on the machine. Liquid's own numbers show the model leading larger MoEs on tool-use and instruction-following tests while still trailing them on knowledge and coding benchmarks.
Who is it for?
developers building local or private agents
Frequently asked questions
- How much does LFM2.5-24B-A2B cost on Liquid AI's API?
- LFM2.5-24B-A2B costs $0.10 per million input tokens ($0.05 when cached) and $1.00 per million output tokens on Liquid AI's API, and Liquid Console gives $2 of starter credit. The weights are also free to download from Hugging Face under Liquid AI's LFM license, so you can run the model on your own hardware instead.
- How does LFM2.5-24B-A2B compare to Qwen3.6-35B-A3B?
- In Liquid AI's own tests LFM2.5-24B-A2B beats Qwen3.6-35B-A3B on AutomationBench (27.69 vs 8.72), LiveClaw (48.09 vs 38.19), ToolSandbox and IFEval. Qwen3.6-35B-A3B stays ahead on knowledge and coding: GPQA Diamond (83.69 vs 74.42), SWE-Bench Verified (66.20 vs 51.00) and WebArena Verified.
- What hardware can run LFM2.5-24B-A2B locally?
- Liquid AI reports LFM2.5-24B-A2B at 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, with up to 21K output tokens per second on one NVIDIA B300 at high concurrency. Only about 2B parameters run per token, but the full 24B model must still fit in memory.
- What changed from LFM2-24B-A2B?
- LFM2.5-24B-A2B keeps the LFM2-24B-A2B architecture but extends context from 32,768 to 131,072 tokens, doubles the vocabulary from 65K to 128K tokens for better non-Latin text, and roughly doubles pre-training from 17T to 33T tokens. Liquid AI also added a five-stage post-training pipeline that ends in agentic reinforcement learning.
Try it
llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF --temp 0.1 --top-k 50 --repeat-penalty 1.05