Liquid AI's LFM2.5 Runs a Full AI Agent on Your Laptop Without Cloud
Liquid AI's LFM2.5-8B-A1B packs a full agentic reasoning loop into 1.5B active parameters, running faster than any rival on consumer CPUs and a single H100.
Takeaways
- New release: Liquid AI's LFM2.5-8B-A1B is an open-weight, on-device MoE model with 8.3B total and 1.5B active parameters, 128K context, trained on 38T tokens.
- Massive benchmark gains: Non-hallucination rate jumped from 7.46 to 63.47; IFEval from 79.44 to 91.84; Tau² Telecom from 13.60 to 88.07 vs. predecessor.
- Speed leader: 253 tokens/s on Apple M5 Max CPU, 28 tokens/s on a Galaxy S26 Ultra phone, and 18.5K tokens/s on a single H100 GPU.
- Reasoning model: First LFM2 edge model to use explicit chain-of-thought, with targeted RL stages to eliminate doom loops and reduce hallucinations.
- Full agentic demo: LocalCowork open-source desktop agent runs 67 tools across 13 MCP servers on a single laptop with no cloud or API keys.
- Limitations: Not recommended for heavy coding tasks or knowledge-intensive QA without retrieval; chain-of-thought adds latency and token cost.
Liquid AI just shipped LFM2.5-8B-A1B, a Mixture-of-Experts model purpose-built to run a complete agentic loop , tool calling, reasoning, and multi-step instruction following , entirely on consumer hardware. No cloud. No API keys. No data leaving the machine. The weights are available now on Hugging Face, and the model is free to use under Liquid's open-weight LFM license.
The number that matters isn't 8B
The model holds 8.3B total parameters but activates only 1.5B per token , that sparsity is what lets it run on consumer hardware. This is the core design bet: a Mixture-of-Experts (MoE) architecture routes each token through only a fraction of the network, keeping compute and memory usage low while preserving the capacity of a much larger model. In Liquid's framing, only 1.5B parameters are active during inference, which changes the budget line from "how big is the model" to "how much of the model wakes up."
Unlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model, producing an explicit chain of thought before its final answer. Liquid adopted this strategy because MoE models generally run in compute-bound settings, where a smaller number of active parameters makes each reasoning token cheap , providing a significant quality boost without compromising speed.
A step change from the previous generation
LFM2.5-8B-A1B builds on LFM2-8B-A1B from October 2025, with an expanded 128K context window, scaled-up pretraining from 12T to 38T tokens, and large-scale reinforcement learning. The benchmark improvements are dramatic across the board:
- A 53.62-point increase on the AA-Omniscience Index and a 56.01-point rise in the non-hallucination rate.
- IFEval (instruction following) rose from 79.44 to 91.84.
- MATH500 climbed from 74.80 to 88.76.
- Tau² Telecom (agentic tool-use benchmark) rose from 13.60 to 88.07.
- IFBench jumped from 26.00 to 56.47 , more than doubling.
The avg@k-based reward , a reinforcement learning technique that scores a model by averaging performance across multiple sampled outputs , enables LFM2.5-8B-A1B to achieve a significantly lower hallucination rate while maintaining reasonable accuracy. It also leads on instruction following benchmarks, matching bigger MoEs like Gemma 4-26B at a fraction of the active parameter count.
How it was built
The architecture has 24 layers: 18 are double-gated LIV convolution blocks and 6 are GQA (grouped-query attention) layers, combining MoE, GQA, and gated short convolution blocks. The hybrid design is what makes long-context inference memory-efficient , the convolution layers don't accumulate a KV cache the way pure attention layers do.
Two training innovations stand out. First, the tokenizer was rebuilt without retraining the model from scratch. Liquid doubled the vocabulary to 128K by extending the existing tokenizer in place, continuing BPE merge training from the original merges on a multilingual corpus, which keeps most existing token IDs as identity mappings and makes every new token decompose deterministically into a sequence of original sub-tokens. The strongest compression gains land in Hindi, Thai, Vietnamese, Indonesian, and Arabic. Thai tokenization efficiency improved by 238% and Hindi by 120%.
Second, Liquid tackled two notorious failure modes of small reasoning models head-on:
- Doom loops , where a model gets stuck repeating fragments of its chain of thought. Liquid added a targeted preference optimization stage that identifies tokens triggering looping behavior and redistributes probability mass toward alternatives. During RL, a lightweight shaping reward also discourages excessive use of common loop-inducing restart words like "Wait…"
- Hallucinations , addressed with a dedicated RL stage using an avg@k-based reward over a diverse knowledge dataset. The goal is to reinforce abstention on queries beyond the model's reliable knowledge, producing a sharper knowledge boundary and clearer expression of uncertainty.
Speed that matches the claims
On CPU, it decodes 253 tokens/s on an M5 Max. When run on a PC equipped with an AMD Ryzen AI Max+ 395, it achieves 148 tokens per second. Even on a Samsung Galaxy S26 Ultra smartphone, a practical output of 28 tokens per second is possible. On the server side, LFM2.5-8B-A1B is the fastest model in its size class, reaching 18.5K output tokens per second at high concurrency , over 1.6B tokens per day on a single H100.
The memory footprint stays under 6 GB, which means it fits comfortably on most modern laptops without any quantization tricks required.
The LocalCowork demo makes it concrete
To show what on-device agentic behavior actually looks like, Liquid released LocalCowork, an open-source desktop agent that now runs on LFM2.5-8B-A1B. The setup is a single laptop, 67 tools across 13 MCP servers, no cloud, no API keys, no data leaving the machine. The dispatch loop , ask, propose, confirm, run, repeat , runs in well under a second per step. The point isn't the individual tools. It's that the full agentic loop feels interactive on hardware you already own.
What it's good at , and where it falls short
Liquid recommends LFM2.5-8B-A1B for agentic workflows, tool use, structured outputs, multilingual assistants, and on-device personal-assistant applications. It is not the best fit for heavy programming or knowledge-intensive question answering without retrieval. The AA-Omniscience accuracy score of 8.67 , while much better than its predecessor , still trails larger models like Qwen3-30B (18.80) on raw factual knowledge. Think of it as a capable agent brain, not an encyclopedia.
LFM2.5-8B-A1B uses explicit chain-of-thought reasoning, which typically improves performance on math and complex reasoning tasks at the cost of higher latency and token usage. For latency-sensitive applications where a fast, direct answer is preferred over a reasoned one, that tradeoff is worth considering.
Getting started
The model is available in multiple formats for every deployment target:
- Transformers , standard
AutoModelForCausalLMload, compatible withtransformers>=5.0.0 - GGUF , for llama.cpp, LM Studio, and Ollama
- MLX , optimized for Apple Silicon
- ONNX , cross-platform, including mobile NPUs
- vLLM / SGLang , high-throughput GPU serving
Here's the minimal Transformers setup:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "LiquidAI/LFM2.5-8B-A1B"
model = AutoModelForCausalLM.from_pretrained(
model_id, device_map="auto", dtype="bfloat16"
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
input_ids = tokenizer.apply_chat_template(
[{"role": "user", "content": "Summarize this document: ..."}],
add_generation_prompt=True,
return_tensors="pt",
).to(model.device)
output = model.generate(
input_ids,
temperature=0.2,
top_k=80,
repetition_penalty=1.05,
max_new_tokens=8192,
)
Fine-tuning is supported via Unsloth and TRL, with Colab notebooks for SFT, DPO, and GRPO provided in the official docs. The model can be customized on a single GPU for any specialized task.
The bigger picture
The broader race is bigger: Gemma, Phi, Llama, Qwen, SmolLM, Granite, Nemotron, Apple Silicon, Copilot+ PCs, Android AICore, Jetson Thor, and DGX Spark are all converging on local inference as a first-class platform. LFM2.5-8B-A1B is Liquid's answer to that race at the 8B scale , the sweet spot where capability and on-device feasibility intersect.
Small models are no longer just fallback models. They are becoming product infrastructure. The assumption that capable agents require cloud infrastructure is being directly challenged. As agentic harnesses become the main way to consume models, LFM2.5-8B-A1B is a first step towards powering on-device, fully private agents. The question for the field is no longer whether a small model can reason and call tools , it's whether you can afford not to run one locally.