DeepSeek AI's model card for DeepSeek V4.1 Flash reports a global KV cache footprint of roughly 890 bytes per token, about a quarter of the 4x-larger footprint of predecessor DeepSeek V4 Flash and a 4…
OpenAI released GPT Image 2.5, an updated image generation model available in two variants — Flare, a faster, lighter version, and Sunburst, a heavier model for demanding generations — accessible thro…
Next AGI released Nex-N2.5 Mini, a multimodal agentic model built on a Qwen3.5 mixture-of-experts backbone, under an Apache 2.0 license on Hugging Face, requiring two 80GB-class GPUs to run. In hands-…
Nex AGI's Nex-N2.5 Mini, the smaller model in the company's N2.5 agentic family, requires two 80GB-class GPUs such as H100s and consumed roughly 66GB of VRAM per card in testing on RunPod, according t…
Two Foundry released TrueForge, an MIT-licensed, model-agnostic open-source agent harness that adds sandboxing, code mode, human-in-the-loop approval gates, and context management to production AI age…
GPT-6 Astra, the latest model variant used inside Codex, shows a sizable jump in 3D spatial reasoning that lets creators generate 3D game environments, characters, and mechanics in Unreal Engine 5 fro…
OpenAI released GPT-6 Astra, a new language model positioned as a major step up from GPT-5.6 Soul, with hands-on testers reporting the largest measurable jump in 3D spatial reasoning and describing th…
Andrej Karpathy's spec-driven method for AI coding reframes work with models like Claude as a three-layer system: a spec, a verifier, and a persistent environment. The spec layer has Claude interview …
OpenAI's Codex coding agent combined with the open-source Hyperframes repo lets users describe a video edit in natural language and receive a finished cut with cropped face cam, animated title cards, …
A practical guide explains how to configure OpenAI's GPT-6 Astra with the Alpaca brokerage API and scheduled tasks to create a 24/7 automated stock trading bot. The setup uses a fixed daily schedule, …
An agent harness is the code and structure around a language model that converts raw text prediction into goal-directed behavior, and it has driven larger performance gains than model upgrades, accord…
OpenAI's GPT-6 Astra, a long-running agentic AI model, can execute multi-step administrative tasks across websites and tools without step-by-step prompting, as demonstrated in a hands-on test involvin…
A new technique called the manager loop lets one AI agent interview a user and delegate tasks to execution agents, enabling supervision of multi-day projects. The approach, which mirrors how people wo…
OpenBMB released MiniCPM5-2B, a 2.5-billion-parameter dense language model designed for on-device deployment, which scores an average benchmark of 53.9, outperforming larger 4B-class models like Qwen3…
ModelBest's MiniCPM5-2B, a dense 2-billion-parameter language model from the Beijing-based AI company spun out of Tsinghua University's NLP lab, claims 2B-class open-source SOTA with an average score …
OpenBMB's MiniCPM5-2B GGUF quantization, tested hands-on with llama.cpp, shows degraded performance on complex tasks: the Q8_0 model invented a nonexistent column in a SQL debugging task and failed a …
Omarchy's new Agent Usage Panel in its top bar shows coding agent plan type, percentage of session and weekly limits consumed, remaining prepaid balance, and token usage broken down by day and model, …
Omarchy, a Linux distribution, uses symlinks to share a single skills folder across multiple AI agents, including Claude Code, OpenAI Codex, and Google Antigravity, pointing five different paths to on…
Prime Agent, a self-improving harness cited at a YC Paper Club talk, pushed ARC-AGI benchmark scores past 95% using the same class of underlying model that previously scored around 30% with earlier ha…
ISTA's Das Lab compressed Qwen3.8-27B, a 27 billion parameter model, to 11.8GB using new GSQ and RCO quantization methods, claiming the compressed model is 'task lossless' and performs close to the fu…