Summary: Liquid AI shipped LFM2.5-8B-A1B on May 28, 2026 as an 8B sparse MoE with about 1B active parameters per token for on-device assistants. Pretraining grew from 12T to 38T tokens and context from 32,768 to 128K, while scores improved sharply on reasoning and tool-calling tests. Its custom LFM 1.0 license—not Apache 2.0—requires review before commercial use.
Liquid AI put out LFM2.5-8B-A1B on May 28, 2026, shipping a Base variant alongside it, and the new model steps in for LFM2-8B-A1B. Under the hood sits a sparse mixture-of-experts design: 8B parameters in total, with roughly 1B active for any given token. The job assigned to it is that of an on-device personal assistant — living on phones, laptops, and edge boxes, organized around tool calling and agentic workflows, and always writing out an explicit chain of thought before it answers.
Size barely moved in this revision; training volume is where the effort went. Pretraining climbed from 12T tokens to 38T, and the context window stretched from 32,768 to 128K. For anyone building local AI, that means the distance between a model on your hardware and one in the cloud keeps shrinking by a different route — more data and a more deliberate training pipeline, not more parameters.
Three Training Stages Take Pretraining From 12T to 38T
Training runs in three stages. It opens with 38T tokens of pretraining. A 2T-token midtraining pass follows, concentrated on reasoning, math, tool use, and long documents, and it lifts the context window to 32K. The final stage contributes 400B further tokens and raises the RoPE base θ, carrying the window the rest of the way to 128K. The vocabulary grew in parallel, from 65,536 to 128K BPE tokens, and the expansion was done in place.
Two choices in the reinforcement learning work deserve a closer look. Preference optimization here goes after “doom loops” — the state where a model gets stuck circling in place and repeating itself. And the reward is computed with avg@k: the same question gets sampled multiple times, and how consistently the answers match determines the score, which is what pulls hallucination down.
Scores Climb Sharply on Reasoning and Tool Calling
Measured against the previous-generation LFM2-8B-A1B (old score → new score):
- IFEval: 79.44 → 91.84
- MATH500: 74.80 → 88.76; AIME25: 20.00 → 42.53, with AIME26 at 50.00
- BFCLv3: 45.07 → 64.36; BFCLv4: 25.52 → 48.50
- Tau² Telecom: 13.60 → 88.07; Tau² Retail: 7.02 → 39.82
- Non-Hallucination Rate: 7.46 → 63.47
The steepest climbs land on the tool-calling suites, BFCL and Tau², which squares with the on-device agent positioning — for an edge agent, tool calling is the capability everything hinges on. Thanks to the wider vocabulary, tokenizer efficiency improved across 16 languages, with Thai (+238.2%), Hindi (+120.4%), Vietnamese (+117.9%), and Arabic (+38.8%) leading the way.
Reasoning-Only, and What It Costs
Reasoning is the only mode this model offers: every answer is preceded by an explicit chain of thought written out in full. Math and multi-step tasks pick up the benefit immediately, and hallucination falls along with them. The bill arrives as latency plus heavier token consumption. Putting that on a phone — the device where response speed is judged first — is a bold stand, but Liquid AI has plainly concluded that a considered answer is worth the wait.
Throughput: 253 Tokens per Second Without a GPU
The published inference figures read as follows: 253 tokens per second on an Apple M5 Max CPU; 146 tokens per second on an AMD Ryzen AI Max+ 395 while staying under 6GB of memory; about 30 tokens per second on phones. Bring in a GPU, and a single H100 SXM5 sustains 18.5K output tokens per second under high concurrency — more than 1.6 billion tokens over a day.
The hybrid architecture carries over: sparse MoE working alongside GQA and gated short convolution blocks. The Hugging Face model card counts 24 layers — 18 convolution, 6 GQA — plus 8.3B total parameters and 1.5B active per token.
One Licensing Catch to Check First
The weights are live on Hugging Face (LiquidAI/LFM2.5-8B-A1B) in GGUF, ONNX, and MLX formats, and the company blog’s phrasing is “download, fine-tune, and deploy without restrictions.” The catch: the model card names the license as LFM 1.0, Liquid’s custom terms — not a standard open-source license like Apache 2.0. Anyone heading for a commercial product should read those terms first.
What Edge AI Builders Should Take Away
Three points stand out. A phone-class model now carries 128K context, which turns long documents, multi-turn conversations, and fully local RAG from theoretical into practical. Tool calling at BFCLv4 48.50 is strong enough to build a local agent around, and the deployment path is covered from day one across LEAP (iOS/Android), llama.cpp, MLX, vLLM, SGLang, and ONNX. Finally, a MoE that activates only about 1B per token adds to the evidence that active parameter count, not total size, decides the on-device experience — welcome news for hardware sizing, memory budgets, and battery life alike.