Tag
OpenBMB has open-sourced MiniCPM5-2B, a 2B-parameter AI model optimized for reasoning, coding, and tool use on resource-constrained hardware, achieving state-of-the-art performance in its size class and demonstrating effective local deployment capabilities.
Release of MiniCPM5-2B-GGUF, a 2B parameter AI model optimized for on-device deployment, achieving state-of-the-art performance in its class with open-source training datasets.
This article provides a detailed test of the local deployment and performance of the Ling-3.0-tiny model on an Apple M5 chip Mac, demonstrating the feasibility of running a 7.9B parameter model at 47 tokens per second without a discrete GPU.
A developer created a 250M parameter quantized LLM from scratch, trained on 30B tokens, deployable in 60 MB on CPU with a novel disk-based long context system supporting up to 100M tokens.
This paper introduces Daedalus-150M, a hybrid language model combining convolution and attention mechanisms optimized for CPU inference, achieving better benchmark performance than larger models with significantly less training data.
This paper addresses deployment bugs in the Nanbeige4.2-3B Looped Transformer model on Apple Silicon and introduces a memory-efficient chunked-prefill strategy to enable agentic tasks.
Cactus Compute releases Needle 2, a 45M-parameter agentic LLM compressed to a 14MB binary for phones, wearables, smart home and robots, achieving 500+ tokens/sec on a Raspberry Pi 5 and running in 28MB RAM.
Liquid AI releases LFM2.5-2.6B, a compact agentic model designed to run entirely on-device, enabling free inference, low latency, and privacy. The post details its training pipeline including SFT, teacher specialization, distillation, and agentic RL.
Liquid AI releases LFM2.5-2.6B, a compact agentic model designed for on-device deployment, supporting tool calling and multi-step workflows with efficient inference on CPUs and GPUs.
SupraLabs releases Supra2-100M Base and Instruct models, a new small language model family with community-driven improvements, benchmarks, and a GGUF version.
Presents B1ade, a minimalist RAG architecture with a 335M zero-training embedding model and a 1B SLM trained via GRPO on 723M tokens, showing emergent attribution behavior and competitive QA performance without large-scale pretraining.
Nanbeige4.2-3B is a compact 3B parameter general agentic model pretrained from scratch with a Looped Transformer, achieving strong agentic and reasoning performance, outperforming larger models on diverse benchmarks. The model and code are open-sourced.
Compares two new AI models for agentic workloads: the compact Nanbeige4.2-3B with a looped transformer architecture and the large Mixture-of-Experts Laguna S2.1, both released on Hugging Face.
OpenLanguageModel (OLM) is an open-source PyTorch library for building and pretraining small language models with readable, composable architecture code, bridging education and research.
OpenBMB released MiniCPM5-2B, a 2 billion parameter language model, currently not yet available on Hugging Face.
This paper investigates whether GRPO post-training improves a small (4B-8B) language and vision-language model web agent. It finds a controlled null result: no configuration yields credible gains on mastered tasks, and moderate-to-high learning rates cause degradation or collapse, revealing a double dissociation between degrade and collapse regimes.
Bilibili releases Index-1.9B, a series of open small language models pre-trained on 2.8 trillion tokens, achieving competitive performance on benchmarks. The four models include base, pure (no instruction data), chat, and a character model with retrieval-augmented generation for role-playing.
A 270M parameter language model trained from scratch on English Wikipedia and instruction-tuned for conversational AI, developed as an independent research project.
Wiola is a novel Small Language Model architecture introducing five independently designed components—SRPE, GCLA, ATM, DSFF, and WiolaRMSNorm—aimed at improving efficiency and coherence, released in sizes from 120M to 1.5B parameters and integrated with HuggingFace Transformers.
Details an approach to train a small LLM judge for evaluating agent outputs, replacing costly frontier models, with a Claude Code plugin for deployment.