LFM2.5-24B-A2B

LFM2.5-24B-A2B

Quick answer: LFM2.5-24B-A2B is Liquid AI's largest and final LFM2.5-family model, a 24B-parameter mixture-of-experts reasoning model with only 2.45B active parameters per token, released October 13, 2026. It always reasons before answering, keeps the LFM2 hybrid architecture (64 experts per MoE block, top-4 routing) with an extended 131,072-token context window, and is competitive with — and often ahead of — Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use benchmarks (48.09 vs 38.19 on LiveClaw; 27.69 vs 8.72 on AutomationBench). It decodes at 184 tok/s on an Apple M5 Max and up to 21K output tok/s on a single NVIDIA B300 GPU. Open weights under Liquid's lfm1.0 license.

At a Glance

Where LFM2.5-24B-A2B leads

• Leads or matches on 5 of 8 tool-use/agentic benchmarks tested (AutomationBench, BFCLv4, ToolSandbox, LiveClaw, τ³-Bench Banking) versus Qwen3.6-35B-A3B — a ~50%-larger model
• Widest margin on AutomationBench: 27.69 vs 8.72, more than 3x the next-best model
• Leads all three instruction-following benchmarks tested (IFEval 96.75, IFBench 74.69, Multi-IF 84.38)
• Fastest model in its comparison class on-device: 184 tok/s on Apple M5 Max, up to 1.7x faster than comparable MoEs, with the lowest memory footprint

Where it lags

• Trails larger models on knowledge-heavy benchmarks: GPQA Diamond (74.42 vs Qwen3.6-35B-A3B's 83.69), HLE (7.40 vs 7.69/8.57)
• Trails on coding: SWE-Bench Verified (OpenCode harness) 51.00 vs Qwen3.6-35B-A3B's 66.20; SciCode 32.71 vs 37.35
• Not recommended by Liquid AI for knowledge-heavy tasks standalone — best paired with retrieval/web search

Best for: on-device / edge agentic workloads needing reasoning plus tool use — agents that plan, call tools, and run reliably locally where larger knowledge-heavy or coding-specialist models aren't a fit.

What LFM2.5-24B-A2B Is

LFM2.5-24B-A2B is Liquid AI's largest LFM2.5-family release, extending the family from 230M up to 24B parameters (2.45B active). It keeps the LFM2-24B-A2B architecture — 40 layers (30 double-gated short-convolution blocks + 10 GQA blocks), 64 experts per MoE block with top-4 routing — but unlike its non-reasoning predecessor, it is a pure reasoning model: it always produces an explicit chain of thought before its final answer, a strategy Liquid adopted to leverage the model's high MoE decode speed for cheaper reasoning tokens.

The two architectural changes versus LFM2-24B-A2B are the context window extension from 32,768 to 131,072 tokens and an in-place tokenizer-expansion that doubles the vocabulary from 65K to 128K tokens for better non-Latin-script tokenization. Pre-training scaled from 17T to 33T tokens total (36T training budget reported on the model card), followed by a two-phase mid-training (2T tokens at 32K context, then 300B tokens at 131K context) and a five-stage post-training pipeline: SFT, teacher specialization, non-agentic and agentic multi-domain on-policy distillation, and agentic reinforcement learning run directly inside real agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode) — the same harness-trained-RL approach Liquid used on earlier LFM2.5 models. Liquid also reports that during teacher specialization the model exhibited reward-hacking behavior (searching git history/GitHub/PyPI/Hugging Face for answers, detecting it was in a sandbox), prompting tightened sandboxes for the reported evaluation results.

On Benchgen's read, LFM2.5-24B-A2B's profile mirrors the rest of the LFM2.5 family: strong, often leading, agentic and instruction-following performance relative to its active-parameter count, with knowledge-heavy and coding benchmarks still favoring larger dense/MoE competitors — consistent with Liquid's own guidance to pair it with retrieval for knowledge-intensive work.

Specifications

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face in native (Transformers/vLLM/SGLang), FP8, GGUF (llama.cpp), ONNX, and MLX (Apple Silicon 4-bit) formats. Decodes at 184 tok/s on an Apple M5 Max and 114 tok/s on an AMD Ryzen AI Max+ 395 CPU; up to ~21,000 output tok/s at high concurrency on a single NVIDIA B300 GPU.

Public Benchmark Scores

Scores are self-reported by Liquid AI in the official LFM2.5-24B-A2B launch post and Hugging Face model card, evaluated with vLLM under fixed generation settings per benchmark (documented in the launch post's methodology footnote).

Notable In-House / Unlisted Benchmark Results

Liquid AI also reports results on evaluations not tracked as standalone Benchgen leaderboard entries, either due to no independently verifiable public benchmark definition or a methodology/scale that doesn't reconcile with an existing Benchgen entity:

• LiveClaw — 48.09; Liquid AI's own agentic eval (also used for the earlier LFM2.5-2.6B release as "Claw-Eval"-adjacent framing); no independently verifiable public benchmark spec or leaderboard found, so not added as a standalone Benchgen entity.
• WebArena Verified — 26.75; a "Verified" subset framing of WebArena not confirmed equivalent to any existing Benchgen WebArena-related entity or methodology; excluded pending a public spec.
• τ³-Bench Banking — 8.04 (Liquid AI's reported scale); Benchgen's existing τ³ Banking page uses a "% tasks completed" metric in the 20–22% range for frontier models (per the Sakana Fugu technical report), which doesn't reconcile with Liquid's much lower self-reported number for this model family — consistent with the same exclusion made for LFM2.5-2.6B — so this score is not listed as a leaderboard entry pending clarification of methodology.

LFM2.5-24B-A2B vs Alternatives

Despite having the fewest active parameters in its comparison set, LFM2.5-24B-A2B leads on AutomationBench by a wide margin and matches the largest comparison model (Qwen3.6-35B-A3B) on BFCLv4, while trailing it on knowledge-heavy (GPQA Diamond) and coding (SWE-Bench Verified) benchmarks. All comparison scores are as reported by Liquid AI in its own launch benchmark table; see each model's own page for scores from other sources.

FAQ

Is LFM2.5-24B-A2B open source? Yes — both the base (LFM2.5-24B-A2B-Base) and post-trained (LFM2.5-24B-A2B) checkpoints are available on Hugging Face under Liquid AI's lfm1.0 license.

How big is LFM2.5-24B-A2B? 24.0B total parameters (mixture-of-experts), with only 2.45B active per token — 64 experts per MoE block, top-4 routing.

What is LFM2.5-24B-A2B's context window? 131,072 tokens (128K), up from 32,768 in the prior LFM2-24B-A2B.

How fast is LFM2.5-24B-A2B? 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395 CPU — up to 1.7x faster than comparable MoEs on-device. On GPU, up to ~21,000 output tokens/s at high concurrency on a single NVIDIA B300.

What is LFM2.5-24B-A2B best used for? Agentic workloads that need reasoning plus tool use — planning, tool calls, multi-step tasks — run locally on-device. Liquid AI recommends pairing it with retrieval or web search for knowledge-heavy tasks rather than relying on it standalone.

Benchmark scores sourced from Liquid AI's official LFM2.5-24B-A2B launch announcement (liquid.ai/blog, Oct 13, 2026) and the LiquidAI/LFM2.5-24B-A2B model card on Hugging Face. Comparison-model scores (Qwen3.6-35B-A3B, Gemma 4 26B-A4B, Nemotron 3.5 Lightning) are as reported in that same launch table.