ReleasesNSFW

Liquid AI's LFM2.5-8B-A1B runs on phones with 1.5B active parameters

Liquid AI released LFM2.5-8B-A1B, an 8-billion-parameter mixture-of-experts model with 1.5 billion active parameters per forward pass, designed for on-device deployment on phones, laptops, and robotics hardware.

ByAlex Sokoloff·May 30, 2026

Liquid AI's LFM2.5-8B-A1B runs on phones with 1.5B active parameters

Liquid AI released LFM2.5-8B-A1B this week, an 8-billion-parameter mixture-of-experts model that activates only 1.5 billion parameters per forward pass. Built for on-device inference on phones, laptops, and embedded systems, the model ships with 128,000-token context, was pretrained on 38 trillion tokens, and underwent large-scale reinforcement learning tuning. According to Liquid AI, the architecture competes with models three to four times its size on tool calling and reasoning tasks.

The weights are available on HuggingFace under the LFM2 open-weight license, which permits commercial use and derivative fine-tunes. An abliterated GGUF quantization also appeared on the platform, tagged as uncensored and ready for imatrix quantization workflows. The open-weight license makes the model a direct competitor to closed on-device APIs and heavier open-weight alternatives that require datacenter GPUs.

More in Releases

Architecture and deployment

LFM2.5-8B-A1B uses Liquid AI's hybrid MoE design from the LFM2.5 family. The 8-billion total parameter count includes multiple expert modules; during inference, the router activates 1.5 billion parameters, keeping memory and compute overhead low enough for real-time use on ARM-based phones and edge devices. The 38-trillion-token pretraining corpus and subsequent RL phase were designed to handle multi-turn tool-use chains. Liquid AI published technical documentation and a public playground for testing alongside the weights.

Early community focus centers on how the 1.5-billion active parameter count performs on extended tool-calling sequences compared to dense 7B and 13B models. The model targets practitioners who need to fine-tune for narrow tasks on a single GPU or deploy inference on consumer hardware without cloud dependencies.

IndustryNSFW

Google's AI search leaves brands blind to how they're described

Google I/O confirmed AI-generated answers now dominate search results, upending SEO strategies built around traditional link rankings and leaving most brands without insight into how AI systems describe them.

May 30, 2026

ReleasesNSFWPlatform

Mero-Artemis-31B-v0.3.1-heretic: uncensored 31B multimodal merge on Gemma4

sh0ck0r released Mero-Artemis-31B-v0.3.1-heretic, a 31-billion-parameter uncensored image-text-to-text model built on Gemma4 and available in safetensors format on HuggingFace.

May 30, 2026

PlatformNSFW

YouTube moves AI labels to top of Shorts, auto-detects synthetic content

YouTube is relocating AI disclosure badges to more visible positions on Shorts and long-form videos and will begin automatically identifying AI-generated content, the company announced this week.

May 30, 2026

ReleasesNSFW

Llama 3.2 3B uncensored GGUF quantization debuts for CPU inference

Jadzblaze99 published quantized GGUF weights for an uncensored fine-tune of Meta's Llama 3.2 3B Instruct model, enabling CPU and low-VRAM deployment of a 3-billion-parameter conversational model without safety filters.

May 30, 2026

Platform

Claude Opus 4.8 hits 69.2% on agent coding with self-assessment fixes

Anthropic's Claude Opus 4.8 ships with improved self-assessment, longer autonomous context retention, and a 4.9-point gain over 4.7 on agent coding benchmarks, now ahead of GPT-5.5.

May 30, 2026