LFM2.5-24B-A2B added to BenchGecko · Liquid AI reasoning MoE, 2B active
Liquid AI released LFM2.5-24B-A2B on 2026-10-13. BenchGecko lists it at $0.05/$0.70 per M tokens with a 128,000 token context window.
What landed
LFM2.5-24B-A2B is a reasoning model from Liquid AI, released on 2026-10-13. It is the largest and final model of the LFM2.5 family, which runs from 230M to 24B-A2B, and the successor to LFM2-24B-A2B. It always produces a chain of thought before its answer and is post-trained for tool calling and agents.
Specs
From Liquid AI's launch post and model card:
- Parameters · 24B total, about 2B active per token (2.45B on BenchGecko)
- Architecture · 40 layers, 64 experts per MoE block with top-4 routing, gated short convolutions plus GQA
- Context · 131,072 tokens, up from 32,768 on LFM2-24B-A2B
- Vocabulary · 128K tokens, up from 65K
- Pre-training · 33T tokens, up from 17T
- Languages · 16
- Weights · open under Liquid's own licence, not Apache
Pricing and context on BenchGecko
BenchGecko lists the row at $0.05 input, $0.025 cached input and $0.70 output per M tokens, with a 128,000 token context and 8,192 max completion tokens, from the routed API listing. Liquid's own API lists $0.10 input, $0.05 cached input and $1.00 output per M tokens.
Vendor-reported benchmarks
From Liquid AI's 2026-10-13 post, compared with Qwen3.6-35B-A3B. Vendor-reported, not BenchGecko tests.
- AutomationBench · 27.69 vs 8.72
- LiveClaw · 48.09 vs 38.19
- BFCLv4 · 65.01
- ToolSandbox · 80.71
- IFEval · 96.75
- IFBench · 74.69
- MATH-500 · 96.04
- GPQA Diamond · 74.42 vs 83.69
- SWE-Bench Verified (OpenCode) · 51.00 vs 66.20
Speed claims
Liquid reports 184 tokens per second decode on an Apple M5 Max, 114 tokens per second on an AMD Ryzen AI Max+ 395, and over 21K output tokens per second on a single NVIDIA B300 at high concurrency.
Sources
- LFM2.5-24B-A2B launch post
- LFM2.5-24B-A2B on Hugging Face
- BenchGecko model page: /model/lfm-2-5-24b-a2b
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.