On May 28, 2026, Liquid AI released LFM2.5-8B-A1B and its Base variant: a sparse mixture-of-experts model with 8B total parameters and roughly 1B active per token, succeeding LFM2-8B-A1B. The positioning is explicit — an on-device personal assistant for phones, laptops, and edge boxes, built around tool calling and agentic workflows, and it generates an explicit chain of thought before answering.
The headline of this revision is not size but training volume: pretraining jumped from 12T tokens in the predecessor to 38T, and the context window grew from 32,768 to 128K. For developers building local AI, that means the gap between “runs on my device” and “runs in the cloud” is closing a different way — not by stacking parameters, but by data volume and a more deliberate training pipeline.
From 12T to 38T: A Three-Stage Training Pipeline
Liquid AI split training into three stages. Stage one is 38T tokens of pretraining. Stage two is a 2T-token midtraining phase focused on reasoning, math, tool use, and long documents, pushing context to 32K. Stage three adds another 400B tokens with an increased RoPE base θ, extending context to 128K. The vocabulary also grew from 65,536 to 128K BPE tokens — expanded in place rather than retrained from scratch.
Two details in the reinforcement learning stage stand out. First, preference optimization targeted at “doom loops,” the degenerate behavior where a model spins in place repeating itself. Second, an avg@k-based reward to reduce hallucination: sample the same question multiple times, and answer consistency feeds directly into the reward.
Benchmark Jump: Tool Calling and Reasoning
Against the previous LFM2-8B-A1B (old → new):
- 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 tool-calling suites (BFCL, Tau²) moved the most, which matches the on-device agentic assistant positioning — tool calling is exactly the capability an edge agent lives or dies on. The tokenizer expansion improved efficiency across 16 languages, with the largest gains in Thai (+238.2%), Hindi (+120.4%), Vietnamese (+117.9%), and Arabic (+38.8%).
The Reasoning-Only Trade-off
LFM2.5-8B-A1B is a reasoning-only model: it writes out an explicit chain of thought before answering. The upside is direct — math and multi-step task scores benefit, and the hallucination rate drops hard. The cost is latency and token consumption. On a phone, where users are most sensitive to response speed, this is a bold choice, but Liquid AI has clearly decided that the quality of a considered answer is worth the delay.
Performance: 253 Tokens per Second on CPU
The official inference numbers: 253 tokens/s on the Apple M5 Max CPU; 146 tokens/s on an AMD Ryzen AI Max+ 395 with under 6GB of memory; roughly 30 tokens/s on phones. On the GPU side, a single H100 SXM5 pushes 18.5K output tokens/s at high concurrency — over 1.6 billion tokens per day.
The architecture keeps Liquid AI’s hybrid design: sparse MoE combined with GQA and gated short convolution blocks. The Hugging Face model card lists 24 layers (18 convolution, 6 GQA), 8.3B total parameters, and 1.5B active per token.
Read the License Before You Ship
The company blog says users can “download, fine-tune, and deploy without restrictions,” and the weights are live on Hugging Face (LiquidAI/LFM2.5-8B-A1B) in GGUF, ONNX, and MLX formats. But the HF model card tags the license as LFM 1.0, Liquid’s custom terms — not a standard open-source license like Apache 2.0. Before putting this into a commercial product, read the fine print.
What It Means for On-Device AI Developers
Three observations. First, 128K context has arrived in a phone-class model: long documents, multi-turn conversations, and fully local RAG become practically feasible. Second, tool calling (BFCLv4 at 48.50) is now good enough to take a local agent seriously, and day-one support for LEAP (iOS/Android), llama.cpp, MLX, vLLM, SGLang, and ONNX makes the deployment path unusually complete. Third, a ~1B-active MoE is more evidence that active parameter count, not total size, determines the on-device experience — good news for hardware sizing, memory budgets, and battery life.
Sources
- LFM2.5-8B-A1B: An Even Better On-Device Mixture-of-Experts — Liquid AI
- LiquidAI/LFM2.5-8B-A1B — Hugging Face
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.