Liquid AI's LFM2.5-24B-A2B: A 24B Open-Weight Agent Model That Runs at Full Speed on a Laptop

The tension between capability and deployability is one of the defining design constraints in modern AI. Liquid AI, the MIT spin-off known for its non-Transformer "Liquid Foundation Models," stepped into that tension on October 13, 2026, with the release of LFM2.5-24B-A2B - a model that squeezes frontier-competitive agentic performance into hardware that fits under a desk.

What It Is

LFM2.5-24B-A2B is a 24-billion-parameter reasoning model, but the number that matters most is smaller: 2.45 billion active parameters per token 1. That split is achieved through a Mixture-of-Experts (MoE) design with 64 experts per block and top-4 routing - only four experts fire for any given token. The result is a model with a large weight footprint but the computational cost of something far smaller 1.

The model is explicitly purpose-built for agents, not chat. It always produces a chain-of-thought before answering, natively calls tools, and handles multi-step tasks inside popular agent harnesses 2. Both the base checkpoint and the post-trained version are open-weight and available for download on Hugging Face today 2.

Architecture and Training

LFM2.5-24B-A2B inherits the skeleton of its predecessor, LFM2-24B-A2B: 40 layers combining 30 double-gated short convolution blocks with 10 grouped-query attention (GQA) blocks 3. Two key upgrades define the ".5" jump:

• Context window: expanded from 32,768 to 131,072 tokens 3
• Vocabulary: doubled from 65K to 128,000 tokens for better non-Latin tokenization 3

Pre-training scaled from roughly 17 trillion tokens to ~33 trillion tokens 4. Mid-training then ran two phases: 2 trillion tokens at 32K context weighted toward reasoning, math, code, and multilingual text, followed by 300 billion tokens at the full 131K context window 4. Tool calls and multi-turn trajectories were woven into both phases to prime the model for agentic workloads 4.

Post-training went deeper still - a five-stage pipeline of supervised fine-tuning, teacher specialization, non-agentic and agentic multi-teacher on-policy distillation, and finally agentic reinforcement learning run inside real production harnesses (OpenClaw, Hermes Agent, Pi, and OpenCode) 5. That last detail is what separates LFM2.5-24B-A2B from a generalist model that happens to support tool-calling syntax: it was trained inside the same frameworks developers will run in production.

Benchmark Results

Liquid AI evaluated the model across 16 benchmarks spanning tool use, knowledge, instruction following, math, and coding, comparing it against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning 6.

On five of the eight agentic and tool-use benchmarks, LFM2.5-24B-A2B leads or matches a field of rivals that carry roughly 50% more total and active parameters 6:

The AutomationBench gap is striking - more than triple the next-best score 6. Instruction following tells a similar story: the model leads all three benchmarks tested (IFEval, IFBench, and Multi-IF) 7.

The trade-offs are real too. On SWE-Bench Verified, LFM2.5-24B-A2B scores 51.00 against Qwen's 66.20, and it trails on GPQA Diamond (74.42 vs 83.69) 8. For hard general reasoning and complex coding tasks, larger cloud models still hold an edge. Liquid AI itself recommends pairing the model with retrieval or web search for knowledge-heavy workloads 8.

Speed: The Headline Number

The reason this matters for on-device deployment is raw throughput. Liquid AI reports 9:

• 184 tokens/s on an Apple M5 Max
• 114 tokens/s on an AMD Ryzen AI Max+ 395
• 21,000+ tokens/s on a single NVIDIA B300 GPU under SGLang 0.5.22

At 184 tokens/s on Apple Silicon, continuous agentic loops - think: fetch tool result, reason, call next tool, repeat - become genuinely interactive at consumer hardware speeds. That throughput figure is reported to be up to 1.7× faster than the next-best comparable model while using the least memory in its class 9.

Deployment Formats

The release ships day-one across every major inference backend 10:

• llama.cpp / GGUF - CPU and local GPU, recommended Q4_K_M quantization
• MLX 4-bit - Apple Silicon via the MLX framework
• vLLM / SGLang - high-throughput GPU serving
• ONNX - cross-platform including cloud, edge, and mobile
• FP8 checkpoint - lower-memory GPU serving with block-wise FP8 weights

An OpenAI-compatible API endpoint is exposed by default, so dropping LFM2.5-24B-A2B into an existing agent stack that calls an OpenAI-compatible server requires no code changes - just a different base URL.

Why It Matters

Most capable open-weight models today are either fast or agent-tuned or runnable on consumer hardware - rarely all three at once. LFM2.5-24B-A2B makes a credible claim at hitting all three targets simultaneously, completing the LFM2.5 family, which now spans two orders of magnitude from the 230M parameter entry to this 24B flagship 10.

For engineers building agentic systems, the implications are concrete: a model that runs privately on-device, at interactive speeds, with competitive tool-use accuracy means cloud API costs and latency are no longer the only viable path to a capable production agent. That's a meaningful shift in the design space - and with open weights now publicly available on Hugging Face, independent verification can begin immediately.