Liquid AI releases LFM2.5-24B-A2B: open-weight 24B reasoning MoE with 2B active parameters and 131K context

Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model with 24B total parameters and 2B active per token, a 131K-token context window, and tool-calling aimed at on-device agents. The weights for the model and its base variant, LFM2.5-24B-A2B-Base, are available on Hugging Face as of October 13, 2026.

Architecture and training

LFM2.5-24B-A2B keeps the architecture of LFM2-24B-A2B: 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention (GQA). What changed:

• Context window: up from 32K to 131K tokens.
• Vocabulary: expanded in place from 65K to 128K tokens, which Liquid AI says improves non-Latin tokenization.
• Reasoning-only: unlike its predecessor, it always produces an explicit chain of thought before answering.
• Pre-training: scaled from ~17T to ~33T tokens. Mid-training adds 2T tokens at 32K context and 300B tokens at 131K context.

Post-training has five stages: supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation (MOPD), agentic MOPD, and agentic reinforcement learning. The RL stage runs inside agent harnesses including OpenClaw, Hermes Agent, Pi, and OpenCode.

Benchmarks

Liquid AI evaluated the model against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B) on 16 benchmarks, all run with vLLM. These are the company's own results, not independent verification.

Benchmark	LFM2.5-24B-A2B	Qwen3.6-35B-A3B
LiveClaw	48.09	38.19
AutomationBench	27.69	8.72
BFCLv4	65.01	65.00
ToolSandbox	80.71	75.02
IFEval	96.75	93.06
IFBench	74.69	64.21
GPQA Diamond	74.42	83.69
SWE-Bench Verified (OpenCode)	51.00	66.20
WebArena Verified	26.75	37.62
PinchBench	83.18	89.93

The model leads or matches on five of eight agentic and tool-use benchmarks (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, τ³-Bench Banking) and on all three instruction-following tests (IFEval, IFBench, Multi-IF at 84.38). On AutomationBench, Liquid AI says it scores more than three times the next-best model.

It trails on several tests. Qwen3.6-35B-A3B leads on GPQA Diamond, PinchBench, WebArena Verified, SciCode (37.35 vs 32.71), and SWE-Bench Verified (OpenCode). Nemotron-3.5-Lightning scores 53.40 on SWE-Bench Verified (OpenCode) and -3.58 on AA-Omniscience, versus -24.90 for LFM2.5. Liquid AI itself recommends pairing the model with retrieval or web search for knowledge-heavy work.

Inference speed

Liquid AI reports the following throughput:

• Apple M5 Max: 184 tokens/s decode.
• Ryzen AI Max+ 395: 114 tokens/s decode.
• Single NVIDIA B300: over 21K output tokens/s at high concurrency.

The company claims the edge figures are up to 1.7x faster than the next-best compared MoE, with the lowest memory use. Day-one support covers llama.cpp, MLX, vLLM, SGLang, and ONNX. The model runs with transformers>=5.0.0 and works behind any OpenAI-compatible backend, so agent harnesses such as Hermes Agent can point at a local endpoint.

Availability and pricing

Both checkpoints are open-weight on Hugging Face. Hosted API pricing is not disclosed, and the source post does not state the license terms or a training data cutoff.

What this means

The 2B active-parameter count is the main point. Reasoning models generate many tokens per answer, so per-token cost decides whether an agent can run continuously on a laptop. Liquid AI's reported 184 tok/s on an M5 Max suggests local reasoning agents are becoming practical on consumer hardware.

The benchmark profile is uneven. Strong tool-use and instruction-following scores fit the agent use case, but the gaps on GPQA Diamond and SWE-Bench Verified (OpenCode) show that active parameter count still limits knowledge and coding depth. Several of the agentic benchmarks (LiveClaw, AutomationBench, PinchBench) are newer and less established, so independent replication will matter. The model is a candidate for private, tool-calling agent workloads, not a general replacement for larger models on coding or knowledge tasks.