Very small and sparse models are moving AI into CPU-only, browser, edge, and mini PC environments. Examples range from 19,969-parameter storytelling and 10M decision models to hybrid 32B MoE and Kolibri-1's 78B total/3.46B active architecture.
The LiquidAI/d1-3B page on Hugging Face, shared without technical details, places the project in the compact LLM segment. The 3B label suggests a model aimed at local inference, self-hosted setups, and data sovereignty constraints. The article analys...
A 19,969-parameter, 81 KB FP32 model generates short stories with goal, problem, and resolution while running on CPU. MacroStories lowers the bound for coherent narrative generation and questions assumptions about local inference.
EmbeddingGemma 2 maps text and images into a 768-dimensional space. Ported to ruNNtime, a TypeScript WebGPU inference library, it enables photo search by description with inference entirely on the browser GPU. No upload, data stays on device: a concr...
TinyDecide is a 10-million-parameter model with a roughly 6 MB footprint, smaller than every entry on the Decision Index leaderboard. It runs in the browser, Node.js, Python, Rust, and on an ESP32. For edge and on-premise deployments, the signal is n...
Blockway releases a 32B preview model with a hybrid architecture. It combines 54 linear-attention layers, 17 compressed-sparse attention layers, one full-attention layer, and hashed n-gram memory in host RAM. Only 18 of 72 layers keep a KV cache, ena...
A researcher trained Apex-2, a decoder-only MoE with 3.87 billion total parameters and 1.45 billion active per token, from scratch on 86.5 billion tokens while scaling from one to two GH200 with DiLoCo. The base checkpoint reaches HumanEval+ parity w...
Aleph Alpha publishes the Kolibri-1 card: 78 billion total parameters, 3.46 billion active, and a 1 million token context under Apache 2.0. AI-RADAR analysis shows the constraint shifting from compute to memory and KV cache, with direct implications ...
The Kyojin project runs two roughly 300-billion-parameter MoE models on a single 128 GB mini PC with AMD Strix Halo. Measured on a Ryzen AI Max+ 395, GLM-5.3-Flash and MiMo-V2.6-Flash reach up to 650 tokens/s prefill and 26 to 44 tokens/s decode. EXL...
Aleph Alpha releases Kolibri-1, an LLM with 78 billion parameters, 3.46 billion active per token, and up to 1 million tokens of context under Apache 2.0. The combination touches three constraints of self-hosted deployments: memory, compute, and licen...
Microsoft has released a 4-billion-parameter model derived from Qwen3.5-4B, trained with reinforcement learning on synthetic SWE environments for agentic software engineering. The approach without behavioral distillation aims to reduce dependence on ...
