lfm ai-models edge-computing developer-tools

Liquid AI reveals 8B-A1B MoE trained on 38T

liquid.ai

May 29, 2026

6 min read

🔥🔥🔥🔥🔥

60/100

Summary

LFM2.5-8B-A1B is an edge model designed for efficient tool calling on consumer hardware, featuring a 128K context window and expanded pretraining from 12 trillion to 38 trillion tokens. The model's vocabulary has been doubled to enhance tokenization for non-Latin languages, enabling it to perform complex tasks on entry-level laptops.

Key Takeaways

LFM2.5-8B-A1B features an expanded context window of 128,000 tokens and a doubled vocabulary size of 128,000 to enhance tokenization efficiency for non-Latin languages.
The model is designed for on-device applications, capable of chaining tool calls and following complex instructions across various devices.
LFM2.5-8B-A1B is a reasoning-only model that produces an explicit chain of thought before arriving at a final answer, improving performance without sacrificing speed.
The model achieves competitive performance with larger models on instruction following and agentic tasks while being the fastest in its size class for CPU and GPU inference.

Read original article

View story →

Coverage & discussions

Hacker News · liquid.ai

Liquid AI reveals 8B-A1B MoE trained on 38T ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Mixed

Positives

Liquid AI's 8B-A1B model demonstrates impressive performance in summarizing long transcripts, showcasing its capabilities despite being a smaller model.
The advancements in smaller models, like Qwen3.5:4B, highlight how effective fine-tuning and reinforcement learning can yield high performance on limited hardware.
Liquid's focus on model training and fine-tuning allows for the creation of specialized tools that are fast, private, and do not require an internet connection.

Concerns

The 8B-A1B model underperformed significantly in bug fixing benchmarks compared to older models, raising concerns about its competitive edge despite being a newer architecture.
There are worries that Liquid AI may be overtraining their models with 38 trillion tokens, which could lead to diminishing returns in performance.