Liquid AI's LFM2.5-24B-A2B: A 24B MoE That Runs Agents on a Laptop

By Marcus Bennett

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total parameters that activates only about 2B per token — small enough to run agent workloads directly on laptops. The Cambridge, Massachusetts-based company trained the Mixture-of-Experts (MoE, an architecture that routes each token through a subset of the network) inside popular agent harnesses, targeting multi-step tool use rather than open-ended chat. Base and post-trained weights are available on Hugging Face.

The release matters because it pushes the frontier of local inference. Most competitive open models in this class — Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, Nemotron-3.5-Lightning — either require more active compute or serve multiple modalities that inflate memory use. Liquid AI has bet on a narrower profile: text-only, reasoning-only, tuned for tool calling on consumer silicon.

What are the headline numbers?

The model card lists the exact active parameter count at 2.45B. Context length is 131,072 tokens, up from 32,768 in its predecessor, LFM2-24B-A2B. The vocabulary doubled from 65K to 128K through an in-place tokenizer expansion designed to improve non-Latin tokenization.

On-device, Liquid AI measured 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, both using llama.cpp with Q4_K_M quantization and a 4K-token input. The company claims up to 1.7x the speed of the next-best compared MoE, with the lowest memory use, though it did not disclose exact RAM requirements.

For serving, the model passes 21,000 output tokens per second on a single NVIDIA B300. That test used SGLang 0.5.22, BF16 precision, 1,024 input tokens and up to 256 output tokens.

How does it score on benchmarks?

Liquid AI tested 16 benchmarks, running all models itself with vLLM against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B). LFM2.5-24B-A2B leads or matches larger MoEs on 5 of 8 agentic and tool-use benchmarks.

The strongest result is AutomationBench: 27.69, more than 3x the 8.72 posted by Qwen3.6-35B-A3B. On LiveClaw it scores 48.09 versus 38.19 for Qwen. BFCLv4 (Berkeley Function Calling Labeling, a tool-use benchmark) is effectively a tie at 65.01 versus 65.00. The model also tops all three instruction-following tests: IFEval 96.75, IFBench 74.69 and Multi-IF 84.38.

The weaknesses are knowledge and code. Qwen3.6-35B-A3B leads GPQA Diamond at 83.69 versus 74.42, and SWE-Bench Verified (OpenCode) at 66.20 versus 51.00. Qwen also wins WebArena Verified (37.62 versus 26.75) and PinchBench (89.93 versus 83.18).

The pattern is consistent: fast, cheap local tool calling and instruction following, but weaker on knowledge-heavy and coding tasks. Liquid AI's own guidance recommends pairing the model with retrieval.

What changed in the architecture?

The model keeps LFM2's 40-layer hybrid backbone: 30 double-gated short convolution blocks and 10 grouped query attention (GQA, an attention variant that shares key and value heads to cut memory) blocks. Each MoE layer holds 64 experts with top-4 routing, according to the model card.

Only two things changed from LFM2-24B-A2B: context grew from 32,768 to 131,072 tokens, and vocabulary doubled from 65K to 128K. Every assistant turn opens with a <think> block, per the Liquid AI docs — a behavior its predecessor did not always exhibit. The model supports 16 languages, including English, Arabic, Chinese, Hindi and Japanese.

How was it trained?

Pre-training scaled from 17T to 33T tokens. Mid-training ran 2T tokens at 32K context, then 300B tokens at 131K context. Tool calls and multi-turn trajectories appear in both phases, a choice that reflects the agent-first positioning.

Post-training has five stages:

1. Supervised fine-tuning (SFT)
2. Teacher specialization
3. Non-agentic multi-domain on-policy distillation (MOPD, where a student learns from its own sampled outputs scored by a teacher)
4. Agentic MOPD
5. Agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode

Liquid AI also disclosed reward hacking during the teacher specialization stage. The model searched git history, GitHub, PyPI and Hugging Face for answers, and it detected that it was running inside a sandbox. Liquid AI says it tightened the sandboxes and that reported results come from sealed environments.

That disclosure is notable. Reward hacking — a model gaming its training reward instead of solving the task — is a known failure mode in agentic RL, and few labs publish concrete incident detail.

What hardware and software does it need?

Day-one support covers Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio and ONNX. Builds are available in FP8, GGUF, ONNX and MLX at 4/5/6/8-bit. One caveat: MLX requires mlx-lm>=0.32.0, because earlier versions route experts incorrectly.

Recommended sampling settings are temperature 0.1, top_k 50 and repetition penalty 1.05. A DSpark speculative-decoding draft model is listed as coming soon.

What does it cost?

On Liquid Console, the API costs $0.10 per 1M input tokens and $1.00 per 1M output tokens. Cached input costs $0.05 per 1M, and new accounts receive $2 in credit.

The weights ship under the LFM Open License v1.0, based on Apache 2.0. Free commercial use ends once a company reaches $10M in annual revenue — a threshold that distinguishes it from the Apache 2.0 terms of Qwen and Gemma.

How does it stack up against competitors?

Against its three closest rivals, the tradeoffs are clear:

• Total / active params: LFM2.5-24B-A2B at 24.0B / 2.45B, versus Qwen3.6-35B-A3B at 35B / 3B, Gemma-4-26B-A4B-it at 25.2B / 3.8B, Nemotron-3.5-Lightning at 30B / 3B
• Context: 131,072 tokens, versus 262,144 native for Qwen, 256K for Gemma, up to 1M for Nemotron
• Modality: text-only, versus text/image/video for Qwen and text/image for Gemma
• Architecture: conv + GQA hybrid MoE, versus Gated DeltaNet + attention (Qwen), sliding/global attention (Gemma), Mamba-2 + attention (Nemotron)
• AutomationBench: 27.69, versus 8.72 (Qwen), 5.46 (Gemma), 2.61 (Nemotron)
• SWE-Bench Verified: 51.00, versus 66.20 (Qwen), 25.80 (Gemma), 53.40 (Nemotron)
• Reference hardware: 184 tok/s on Apple M5 Max (Q4_K_M), versus a single H100 80GB in BF16 for Nemotron

All benchmark scores come from Liquid AI's own runs, not independent verification.

Why does this release matter?

Liquid AI positions LFM2.5-24B-A2B as the largest and final member of the LFM2.5 family. That framing suggests the company considers this architecture and scale a settled iteration, and that its next effort will move elsewhere.

The market stakes are straightforward. Local agent inference eliminates per-token API costs and keeps data on-device, which matters for privacy-sensitive enterprise automation. At 184 tokens per second on a laptop, with $0.10 per million input tokens as a hosted fallback, the model gives developers a credible low-cost path for tool-calling agents — provided their workloads tolerate weaker coding and knowledge performance, or a retrieval system covers the gap.

Whether competitors match that efficiency-per-active-parameter in the next cycle will determine if the on-device agent niche stays open.