Liquid AI's LFM2.5-24B-A2B landed on Hugging Face on October 13 as a successor to LFM2-24B-A2B, and the interesting number is not the 24B in the name — it is the 2B active. Only 2B parameters fire per token, so each reasoning token is cheap enough that you can run an agent loop continuously on hardware you already own instead of paying per call.
What Actually Changed From LFM2-24B-A2B
The architecture carries over: 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions and GQA. The deltas are what matter for agents:
| Spec | LFM2-24B-A2B | LFM2.5-24B-A2B |
|---|---|---|
| Context window | 32K | 131K tokens |
| Vocabulary | 65K | 128K tokens |
| Pre-training tokens | ~17T | ~33T |
| Reasoning mode | — | reasoning-only, explicit chain of thought |
Mid-training adds 2T tokens at 32K context weighted toward reasoning, math, code, instruction following, and multilingual text, plus 300B tokens at 131K context weighted toward long-context data. Tool calls and multi-turn trajectories appear in both phases. The tokenizer expansion is in-place and doubles the vocabulary, which the post says improves non-Latin tokenization.
The Benchmark Table You Should Read Twice
Liquid AI ran 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B), all with vLLM and each benchmark's recommended generation parameters.
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B | Gemma-4-26B-A4B-it | Nemotron-3.5-Lightning |
|---|---|---|---|---|
| AutomationBench | 27.69 | 8.72 | 5.46 | 2.61 |
| ToolSandbox | 80.71 | 75.02 | 72.09 | 75.13 |
| LiveClaw | 48.09 | 38.19 | 41.25 | 37.72 |
| BFCLv4 | 65.01 | 65.00 | 57.09 | 53.78 |
| IFEval | 96.75 | 93.06 | 92.00 | 94.19 |
| IFBench | 74.69 | 64.21 | 47.44 | 71.67 |
| SWE-Bench Verified (OpenCode) | 51.00 | 66.20 | 25.80 | 53.40 |
| GPQA Diamond | 74.42 | 83.69 | 74.90 | 72.42 |
The pattern is consistent with the angle Liquid AI is pushing: tool use and instruction following are where this model earns its keep, and it leads or matches on five of eight agentic benchmarks. On AutomationBench it scores more than three times the next-best model. On knowledge-heavy (GPQA Diamond, HLE, AA-Omniscience) and coding (SWE-Bench Verified, SciCode) benchmarks, larger models keep the edge — SWE-Bench Verified is 51.00 versus Qwen's 66.20. That is the honest caveat: this is not your coding agent replacement, it is your tool-calling agent on your own box.
Inference Speed Where It Matters
The model ships day-one support across llama.cpp, MLX, vLLM, SGLang, and ONNX — that matters more than the headline throughput because it means your existing serving path probably works.
Edge numbers: 184 tok/s on an Apple M5 Max, 114 tok/s on a Ryzen AI Max+ 395. Liquid AI claims up to 1.7x faster than the next-best model in the comparison while using the least memory. On a single NVIDIA B300 it reaches over 21K output tok/s at high concurrency.
Hooking It Into Your Harness
Install the latest transformers (transformers>=5.0.0):
pip install -U transformers
Then load with the standard API — AutoModelForCausalLM.from_pretrained("LiquidAI/LFM2.5-24B-A2B", device_map="auto", dtype="bfloat16"), generate with temperature=0.1, top_k=50, repetition_penalty=1.05.
The interesting bit for agent operators is that the model is designed to be served behind any OpenAI-compatible backend and pointed at by a harness. With Hermes Agent, that is:
hermes config set model.provider custom
hermes config set model.base_url http://localhost:8080/v1
hermes config set model.default LFM2.5-24B-A2B
hermes config set model.context_length 131072
hermes config set model.api_mode chat_completions
hermes config set agent.tool_use_enforcement true
The port depends on your backend: llama.cpp and MLX use 8080, vLLM uses 8000, SGLang uses 30000, LM Studio uses 1234. The model card has OpenClaw and Pi configs.
Where I'd Reach For This
If you are building a feature where data cannot leave the device — a budgeting assistant, a local triage bot, anything touching PII — this is the first model in its weight class that has the tool-calling behaviour to handle agent loops and the throughput to feel instant rather than tolerable. The 131K context means you can keep longer working sets resident without juggling a sliding window.
The catch: keep retrieval or web search wired up for knowledge-heavy questions, because the GPQA and HLE gap versus Qwen3.6 is real. And the fine-tuning path is a tutorial, not a promise — check the model card before you assume LoRA works out of the box.
Both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight and available now.
Sources: