One billion fire, eight billion sit in memory

LFM2.5-8B-A1B holds eight billion parameters and fires about one billion of them on each token. The full weight set sits in memory; the mixture-of-experts router picks a sparse slice to actually compute. That is what lets Liquid AI’s new edge model decode 253 tokens per second on an M5 Max while staying under 6 GB, 146 on a Ryzen AI Max+ 395, and roughly 30 on a phone. The whole model is resident, but the per-token cost tracks the active parameter count, not the total.

That arithmetic drove a design choice: LFM2.5-8B-A1B is reasoning-only. It emits an explicit chain of thought before the final answer. Liquid’s stated reason for making this the default is that MoE inference tends to be compute-bound, so a small active parameter count makes each reasoning token cheap. You get the quality lift from longer deliberation without paying the latency you would on a dense model of the same total size.

What changed since October

The model builds on LFM2-8B-A1B from October 2025. The context window went from 32,768 to 128,000 tokens. Pretraining scaled from 12 trillion to 38 trillion tokens. The vocabulary doubled from 65,536 to 128,000, with the compression wins called out in Hindi, Thai, Vietnamese, Indonesian, and Arabic. The rest of the architecture follows the same combination as before: MoE blocks, grouped-query attention, and gated short convolution blocks.

The tokenizer expansion is the part worth copying. Rather than retrain from scratch on the larger vocabulary, Liquid extended the existing 65K tokenizer in place. They continued 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. New embedding rows were initialized as the mean of their sub-token decompositions; shared rows were copied unchanged. Quality came back through a short two-stage adaptation: embedding-only training, then full-model continued pretraining. The jump to 128K context used the usual lever, raising the RoPE base θ, plus a 400B-token midtraining stage on long-document and long-trajectory data.

Two failure modes they trained against

Small reasoning models have two characteristic problems, and Liquid addressed both explicitly. The first is doom loops: long reasoning traces that get stuck repeating. They added a preference optimization stage that finds the tokens which trigger looping in specific contexts and redistributes probability toward plausible alternatives, leaving the rest of the next-token distribution mostly intact. During RL they also added a shaping reward that discourages overusing restart words like “Wait…”.

The second is hallucination. An eight-billion-parameter model with one billion active has limited knowledge capacity, so it invents facts past its boundary. The fix was an RL stage with an avg@k-based reward over a diverse knowledge dataset, tuned to reinforce abstention on queries beyond reliable knowledge while preserving what the model does know. The claimed result is a lower hallucination rate with accuracy held roughly steady, and clearer expression of uncertainty. On instruction following the model is reported to match Gemma 4-26B at a fraction of the active parameter count, and it is strong on the Tau2-Telecom agentic benchmark.

Running it

Day-one support covers llama.cpp (GGUF), MLX, vLLM, SGLang, ONNX, and Liquid’s own LEAP for iOS and Android. The CPU path is the headline number: under llama.cpp it was the fastest model Liquid tested at both prompt reading and generation on laptop-class chips, under 6 GB throughout. On GPU, measured on a single H100 SXM5 with SGLang 0.5.12 at 1,024 input and up to 256 output tokens in BF16, it reaches 18.5K output tokens per second at high concurrency, over 1.6 billion tokens per day.

The use case these numbers are aimed at is local tool calling. Liquid’s Localcowork demo runs the model against 67 tools across 13 MCP servers on a single laptop, no cloud, no API keys, nothing leaving the machine, with dispatches landing in well under a second each and full audit trails. Both the base and post-trained checkpoints are open-weight and available on Hugging Face.