Overview
LFM2.5-24B-A2B is an open-weight model Liquid AI released on October 13, 2026, and the largest, described by the company as the last, model in the LFM2.5 family. It carries 24 billion total parameters with 2 billion active per token, and the mixture-of-experts design separates how much the model knows from how much work it does per token, which is why a model this size still runs on a single machine.
Against the previous LFM2-24B-A2B the changes land in four places: context grows from 32K to 131K tokens, the vocabulary doubles from 65K to 128K, which the company says helps non-Latin languages, pretraining scales from roughly 17T to 33T tokens, and the model becomes reasoning-only, always writing out a chain of thought before its final answer. Liquid AI argues the extra thinking costs little time because decoding is fast.
Speed is the headline. Liquid AI reports 184 tokens per second on an Apple M5 Max, 114 tokens per second on an AMD Ryzen AI Max+ 395, and over 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. Both the base model and the post-trained model are open weight, available on Hugging Face and in the company Playground, with llama.cpp, MLX, vLLM, SGLang and ONNX supported on day one.
Key Features
- 24B total, 2B active: MoE design with 40 layers, 64 experts per block with top-4 routing, gated short convolutions and grouped-query attention
- 131K context: Up from 32K, with mid-training adding 300B tokens of long-context data
- 128K vocabulary: Doubled from 65K, which the company says improves tokenization for non-Latin languages
- Reasons before answering: A reasoning-only model that always writes an explicit chain of thought first
- 33T token pretraining: Up from roughly 17T, with mid-training stages weighted toward reasoning, math, code, instruction following and multilingual text
- Five-stage post-training: Supervised fine-tuning, teacher specialisation, two rounds of multi-domain on-policy distillation, then agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode
- On-device speed: Company-reported 184 tokens per second on an Apple M5 Max, 114 on a Ryzen AI Max+ 395, over 21,000 output tokens per second on one B300 at high concurrency
- Open weights and tooling: Base and post-trained models are downloadable, with llama.cpp, MLX, vLLM, SGLang and ONNX support on day one
Use Cases
- Running agents on your own or your customer hardware so data stays local
- Long-context tool calling and multi-step work without depending on a cloud API
- Serving the model behind any OpenAI-compatible backend and pointing an agent harness at that endpoint
- Teams working in non-Latin languages who want better tokenization
- Shops already running llama.cpp, vLLM or MLX who want an on-device model that reasons
Pros
- Open weights you can download and run on your own hardware, with no data leaving your machines
- Only 2B parameters active per token, so decoding speed sits at the front of its class on device
- 131K context, so long tasks no longer need to be cut into pieces
- Doubled vocabulary helps non-Latin languages
- Day-one support across llama.cpp, MLX, vLLM, SGLang and ONNX keeps integration cheap
Pricing
The weights are free to download and self-hosting costs whatever your own hardware and operations cost. The Liquid AI Playground and hosted services follow the company published pricing.
Summary
LFM2.5-24B-A2B is the open-weight model Liquid AI released on October 13, 2026 as the closing model of the LFM2.5 family. It pairs 24B total parameters with 2B active per token in a MoE layout, separating knowledge capacity from per-token compute so it fits on one machine. Against the previous generation, context goes from 32K to 131K, vocabulary doubles from 65K to 128K, pretraining grows from roughly 17T to 33T tokens, and the model becomes reasoning-only, writing a chain of thought before answering. Company-reported speed is 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395.
The benchmarks are Liquid AI own numbers. The company says the model leads or matches on five of eight tool-use and agentic benchmarks, including AutomationBench at 27.69, LiveClaw at 48.09, BFCLv4 at 65.01, ToolSandbox at 80.71 and tau3-Bench Banking, and leads all three instruction-following benchmarks with IFEval at 96.75. Larger models keep the lead on knowledge-heavy and coding benchmarks, where it scores 74.42 on GPQA Diamond and 51.00 on SWE-Bench Verified with OpenCode, and the company advice is to pair it with retrieval or web search for that kind of work.
In short: if you want a tool-calling agent that finishes quietly on local hardware, this fits. If you want knowledge-heavy question answering or competitive-level coding, the trade-off at this size runs the other way.
Version History
- LFM2.5-24B-A2B releases as the largest and final model of the LFM2.5 family, built for fast private on-device agents (2026-10-13): On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, which it describes as the largest and final model in the LFM2.5 family. The model pairs 24B total parameters with 2B active per token, keeping the previous architecture of 40 layers with 64 experts per block and top-4 routing. Context grows from 32K to 131K tokens, the vocabulary doubles from 65K to 128K, pretraining scales from roughly 17T to 33T tokens, and the model becomes reasoning-only, emitting an explicit chain of thought before its answer. Liquid AI reports 184 tokens per second on an Apple M5 Max, 114 tokens per second on a Ryzen AI Max+ 395 and over 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. Both the base and post-trained models are open weight, with llama.cpp, MLX, vLLM, SGLang and ONNX supported on day one.