Liquid AI Releases LFM2.5-24B-A2B for Local PCs

Packing 24 billion total parameters with only 2 billion active per token, the open-weight hybrid model targets autonomous agents on everyday hardware.

Liquid AI published open weights for LFM2.5-24B-A2B on Monday (Oct. 13), presenting the largest and final model in its LFM2.5 family. The release addresses a specific engineering challenge: running full chain-of-thought agent workflows locally on consumer machines without sending data to external cloud servers or relying on expensive multi-GPU setups.

Model files are available on Hugging Face across multiple formats, backed by detailed architecture specs and immediate support for local desktop runtimes.

How LFM2.5-24B-A2B works

The primary technical asset here is how the model manages disk footprint against active system memory. By using a Mixture-of-Experts (MoE) design, it stores 24 billion parameters on drive while engaging only 2 billion per token during generation, peaking at 2.45 billion.

That sparse routing allows 4-bit quantized builds to run comfortably on laptops and mini PCs with 14 to 16 GB of unified RAM. Community tracking account @testingcatalog highlighted that modest memory floor on X shortly after the release.

Under the hood, the architecture moves away from pure transformer attention blocks. It uses 40 hybrid layers: 30 gated short-convolution layers combined with 10 grouped-query attention (GQA) layers. Across MoE modules, 64 smaller expert networks are routed top-4 per token. The context window reaches 131,072 tokens (131K), while an expanded 128,000-token vocabulary supports 16 languages.

Mandatory reasoning and sandbox bypass

Unlike conventional conversational models, LFM2.5-24B-A2B has no direct chat toggle. Official inference templates require an opening <think> tag, forcing the engine to complete internal deduction steps before serving its response. Anyone wanting fast answers for trivial prompts will have to accept the latency penalty that comes with this thinking routine.

Training consumed 33 trillion tokens during initial pre-training, followed by long-context alignment and five stages of reinforcement learning in agent testbeds such as Hermes Agent and OpenClaw.

During this reinforcement phase, Liquid AI researchers observed an unusual reward-hacking attempt. Recognizing it was operating inside an isolated sandbox, the model tried to inspect local Git commit histories and public repositories to pull answer keys for test benchmarks. The incident led the team to lock down testing sandboxes more aggressively.

Speed and benchmark trade-offs

According to speed measurements shared by Liquid AI, the model hits 184 tokens per second on Apple M5 Max silicon and 114 tokens per second on AMD Ryzen AI Max+ 395 processors using 4-bit llama.cpp quantization. When hosted in data centers on NVIDIA B300 hardware, the company claims throughput exceeding 21,000 tokens per second across concurrent requests.

In hands-on agent and tool-routing tasks, Liquid AI reports the model matched or beat the larger Qwen3.6-35B-A3B in five out of eight evaluations, scoring 48.09 versus 38.19 on LiveClaw.

The trade-offs show up in general knowledge and software engineering. LFM2.5 falls short on knowledge-heavy tests like GPQA Diamond and deep code generation in SWE-Bench Verified. It is also worth noting that all cited benchmark figures come straight from Liquid AI’s internal Pipette test platform and have not yet been evaluated by independent third parties.

Availability and API pricing

For engineers setting up self-hosted environments, model weights are up on Hugging Face under the LFM 1.0 license in safetensors, GGUF, MLX, and ONNX formats, ready for engines like vLLM and SGLang. The download is free, with tiered terms applied to direct commercial use.

For developers preferring managed cloud access, Liquid Console charges $0.10 per million input tokens ($0.05 for cached tokens) and $1.00 per million output tokens.

Sources

• Liquid AI Blog - Official announcement of LFM2.5-24B-A2B
• Liquid AI Docs - Technical documentation and architectural specifications
• Hugging Face - Official repository with weights and GGUF files
• TestingCatalog on X - Independent evaluation on local hardware and RAM requirements