MiniCPM5-1B’s RL+OPD Training Hits SOTA on Complex Tasks; Granite 4.1 8B Balances Light Params with Enterprise‑Grade Capabilities

MiniCPM5-1B, the first model of the MiniCPM5 series released by OpenBMB, is a 1 B‑parameter dense Transformer designed for edge deployment. It follows the standard Llama architecture, adds a mixed‑reasoning paradigm with a <think> tag, and is trained with RL+OPD, which boosts core performance and removes output redundancy. The model supports a 131 K context window and achieves open‑source SOTA on agentic tool calling, code synthesis and high‑difficulty reasoning tasks.

HiDream‑O1‑Image, introduced by the HiDream.ai team in 2026, is a native unified image generation model built on a pixel‑level Unified Transformer (UiT). Unlike traditional pipelines, it encodes pixels and text in a shared token space without external VAE or a separate text encoder.

X2SAM, released in April 2026 by Sun Yat‑sen University, Peng‑Cheng Laboratory and Meituan, is a multimodal model for unified image and video segmentation. It integrates textual, visual prompts and segmentation into a single interaction flow.

LocateAnything‑3B, launched by NVIDIA in 2026, is a 3 B‑parameter visual‑language localization model in the Eagle VLM series. Its Parallel Box Decoding predicts full bounding‑box coordinates as structured blocks, improving throughput while preserving geometric consistency.

Granite 4.1 8B, an IBM open‑source foundation model released in 2026, is an 8 B‑parameter dense decoder that remains lightweight yet delivers enterprise‑grade performance. It natively supports multilingual tasks, coding, retrieval‑augmented generation, tool usage and structured JSON output.

The newsletter also offers a compute‑gift package containing GPUs such as NVIDIA RTX 5090 and PRO 6000, and provides online demo links for each model (e.g., https://go.hyper.ai/OBlhv for MiniCPM5‑1B). Additional resources include curated tutorials, popular AI encyclopedia entries, and upcoming conference deadlines.