CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
← Back to feed

Liquid AI LFM2.5-24B-A2B: 2B-Active MoE Tops Qwen3.6-35B on 5 of 8 Agent Benchmarks

Liquid AI released LFM2.5-24B-A2B on October 13, the final and largest member of the LFM2.5 model family. It is a 24-billion-parameter sparse Mixture-of-Experts architecture that activates roughly 2 billion parameters per token — keeping on-device inference practical — while carrying a 131,072-token context window.

The model and its base checkpoint are available as open weights on Hugging Face.

Benchmark Comparison Against Qwen3.6-35B-A3B

Liquid’s launch post frames the model against Qwen3.6-35B-A3B, a similarly active-parameter MoE with a 35B total count. LFM2.5-24B-A2B is competitive or better on five of eight agentic benchmarks:

BenchmarkLFM2.5-24B-A2BQwen3.6-35B-A3B
LiveClaw48.0938.19
Automat27.698.72
(3 additional agentic benchmarks)aheadbehind

The LiveClaw margin (48.09 vs 38.19) is a 26% relative advantage. Automat — which evaluates multi-step workflow automation — shows an even starker gap: 27.69 vs 8.72.

Qwen3.6-35B-A3B outperforms on the remaining three of eight benchmarks, making this an architectural trade-off rather than a clean sweep.

What Makes It Different From Earlier LFM2.5 Models

The LFM2.5 family previously topped out at the 8B model, which Liquid shipped in May 2026 with 38 trillion training tokens and a focus on reducing hallucinations. The 24B-A2B is a scale jump: 3x the total parameters, with the active-parameter count kept low enough for consumer GPU inference.

The earlier 2.6B model (August 2026) led every rival up to 10B parameters on tool-use and agentic tasks. The 24B-A2B is designed for tasks that exceed the 2.6B’s capacity — longer context, more complex multi-step plans — while preserving the same on-device deployment story.

Significance

Most frontier agent benchmarks are currently dominated by cloud-hosted models in the 200B+ range. An open-weight 24B MoE that outperforms a 35B rival on over half the tested agentic tasks, and runs on-device, is a practical capability gain for applications that can’t send data to a cloud API — edge devices, privacy-constrained deployments, and latency-critical pipelines.

Liquid has not announced pricing for API access to the 24B-A2B. The model is available for local deployment via Hugging Face.