LFM2.5-8B-A1B: an Even Better on-Device Mixture-of-Experts

Today, we're releasing LFM2.5-8B-A1B, an edge model built for fast, reliable tool calling on consumer hardware.It builds on our LFM2-8B-A1B release from October 2025, with an expanded 128K context window, scaled-up pretraining (from 12T to 38T tokens), and large-scale reinforcement learning. We also doubled its vocabulary to improve tokenization efficiency for non-Latin languages. The result is a model that chains tool calls, achieves tasks, and fits comfortably even on an entry-level laptop.The base (LFM2.5-8B-A1B-Base) and post-trained (LFM2.5-8B-A1B) models are available today on Hugging Face and our Playground. Check out our docs on how to run and fine-tune them locally.*AA-Omniscience Index (higher is better) rewards correct answers and penalizes hallucinations. Scores range from -100 to 100. See more results on Artificial Analysis.HighlightsOn-device personal assistant. Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices.Compressed performance. Competitive with much larger dense and MoE models on instruction following and agentic tasks.Unmatched throughput. Fastest in its size class on both CPU and GPU inference, with day-one support for llama.cpp, MLX, vLLM, and SGLang.What changed since LFM2-8B-A1BCompared to LFM2-8B-A1B, this new version expands the context window from 32,768 to 128,000 tokens. This allows the model to process longer documents and reason for longer. Its vocabulary size was also scaled up from 65,536 to 128,000 to tokenize non-Latin scripts more efficiently. We see particularly strong compression gains in Hindi, Thai, Vietnamese, Indonesian, and Arabic. The rest of the architecture follows the same combination of MoE, GQA, and gated short convolution blocks as LFM2-8B-A1B, as shown in the following figure.

Unlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model, producing an explicit chain of thought before its final answer. We adopted this strategy because MoE models generally run in compute-bound settings, where a smaller number of active parameters makes each reasoning token cheap. This provides a significant quality boost without compromising speed.Thanks to reasoning and scaled-up training, this new version performs significantly better: Benchmark LFM2-8B-A1B LFM2.5-8B-A1B Δ

AA-Omniscience Index -78.42 -24.70 +53.62

AA-Omniscience Accuracy 7.33 8.67 +1.34

AA-Omniscience Non-Hallucination Rate 7.46 63.47 +56.01

IFEval 79.44 91.84 +12.40

IFBench 26.00 56.47 +30.47

Multi-IF 58.54 79.93 +21.39

MATH500 74.80 88.76

+13.96

AIME25 20.00 42.53 +22.53

BFCLv3 45.07 64.36 +19.29

BFCLv4 25.52 48.50 +22.98

Tau² Telecom 13.60 88.07 +74.47

Tau² Retail 7.02 39.82 +32.80

Training highlightsTokenizer expansion. LFM2-8B-A1B was originally trained with a 65K BPE tokenizer optimized for our initial language coverage. To better support non-Latin scripts in LFM2.5, we doubled the vocabulary to 128K by extending the existing tokenizer in place rather than retraining the model from scratch.. We 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. We initialize the new embedding rows as the mean of their sub-token decompositions and copy the shared rows unchanged. We then recover quality through a brief two-stage adaptation: embedding-only training, followed by full-model continued pretraining.The table below reports chars/token, roughly how much text each token carries: higher is better, and the new tokenizer is more efficient in all 16 languages Tokenizer Arabic (ar) German (de) English (en) Spanish (es) French (fr)

Hindi (hi) Indonesian (id) Italian (it) Japanese (ja) Korean (ko) Polish (pl) Portuguese (pt) Russian (ru) Thai (th) Vietnamese (vi) Chinese (zh)

Old tokenizer 2.239 3.641 4.063 3.442 3.618 0.961 2.731 3.251 1.836 1.652 2.672 3.194 2.703 0.671 1.519 1.475

New tokenizer 3.107 3.783 4.137 3.579 3.759 2.118 3.513 3.475 1.963 1.943 2.895 3.450 2.876 2.269 3.311 1.620

Improvement +38.8% +3.9% +1.8% +4.0% +3.9% +120.4% +28.6% +6.9% +6.9% +17.6% +8.3% +8.0% +6.4% +238.2% +117.9% +9.8%

Context extension. We first extended the context window to 32K through a 2T token midtraining phase focused on reasoning, math, tool-use, and longer documents.

We then extended the context to 128K by increasing the RoPE base θ and running an additional 400B token midtraining stage focused on long-document and long-trajectory data.Doom loops. We added a targeted preference optimization stage to reduce doom loops in long reasoning traces. This stage identifies tokens that tend to trigger looping behavior in specific contexts, then redistributes probability mass toward plausible alternatives, while leaving the rest of the next-token distribution largely intact. During RL, we also added a lightweight shaping reward that discourages excessive use of common loop-inducing restart words like “Wait…”. We'll share more details on the full pipeline, objective, and empirical results in a dedicated blog post.Hallucinations. Because of their small number of parameters, edge models have a limited knowledge capacity, which leads to more hallucinations. To mitigate hallucinations, we added a targeted RL stage that uses an avg@k-based reward over a diverse knowledge dataset. The goal is to reinforce abstention on queries beyond reliable knowledge while preserving existing knowledge. This produces a sharper knowledge boundary and clearer expression of uncertainty.BenchmarksWe evaluated LFM2.5-8B-A1B across benchmarks covering knowledge, instruction following, math, and agentic workflows. The model is competitive with both dense alternatives with a similar total number of parameters and much larger MoEs. Model Parameters AA-Omniscience Index Accuracy Non-Hallucination IFEval IFBench Multi-IF

LFM2.5-8B-A1B 8B/A1B -24.70 8.67 63.47

91.84 56.47 79.93

Granite-4.0-H-Tiny 7B/A1B -75.50 9.37 6.38 82.23 21.28 59.00

Qwen3.5-4B 4B -51.53 17.20 16.99 87.80 50.38 67.43

Qwen3-30B-A3B-Thinking-2507 30.5B/3.3B -51.31 18.80 13.87 90.82 51.11 79.04

Gemma-4-E2B-IT 5.1B -72 7.00 15.05 82.93 33.53 69.70

Gemma-4-E4B-IT 8B -50.67 8.10 36.06 87.74 39.48 77.58

Gemma-4-26B-A4B-IT

26B/4B -62.07 14.37 10.75 91.40 47.25 82.06

gpt-oss-20b 21B/3.6B -49.17 14.57 24.50 86.73 58.65 76.64

The avg@k-based reward 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.Math and agentic workflows Model Parameters MATH500 AIME25 AIME26 BFCLv3 BFCLv4 Tau² Telecom Tau² Retail

LFM2.5-8B-A1B 8B/A1B 88.76 42.53 50.00 64.79 49.73 88.07 39.82

Granite-4.0-H-Tiny 7B/A1B 59.20 4.93 3.33 56.89 28.52 16.67

18.42

Qwen3.5-4B 4B 80.76 54.28 58.33 71.06 54.01 87.72 71.93

Qwen3-30B-A3B-Thinking-2507 30.5B/3.3B 86.48 71.67 66.67 73.39 50.53 21.93 56.14

Gemma-4-E2B-IT 5.1B 64.00 26 30 56.44 31.91 22.37 18.95

Gemma-4-E4B-IT 8B 65.00 34.33 40.67 57.31 33.92 26.75 42.11

Gemma-4-26B-A4B-IT 26B/4B 94.20 68.67 72.00 68.87 55.87 42.11 55.26

gpt-oss-20b 21B/3.6B