open-weight

Reflection AI introduced Beam, a 501B-parameter sparse MoE open-weight model (23B active) built for coding, reasoning, and agentic workloads, pretrained on 23.8T tokens and scaled with a high-compute RL run on 10.5K GB300 GPUs. Beam is positioned as advancing the Western open-weight frontier, approaching GLM 5.2 and Qwen 3.8-Max on coding/agentic tasks while using 3–4× less inference compute, with weights and a technical report due later this month.

Upstage's Solar Mini 4, a 3B-active-parameter (35B total) model with 512K context, is available free on Nous Portal for two weeks, with benchmark scores above models having 10x the active parameters.

Sources report that several Western open-weight AI models are expected to launch this month, including Reflection AI's first model, which aims to rival top Chinese open-weight models.

Aleph Alpha released Kolibri, an open-weight sovereign German/English Mixture-of-Experts LLM with 78B total parameters (~3.5B active per token), trained from scratch on German and Finnish infrastructure under Apache 2.0, with weights on Hugging Face and a 262k-token native context.

Runway Research introduces Praxis-1, its first open-weight world action model for robotics, built on large-scale video pretraining to learn generalist policies across any embodiment and environment. The model is being tested with partners like Noble Machines, Standard Bots, and Ultra ahead of a public release in the coming months.

Anthropic's report reveals that the Chinese open-weight AI model GLM-5.3 can build cyber exploits comparable to their Mythos Preview model, with easily bypassed safety guardrails, posing significant cybersecurity risks.

Anthropic's research analysis finds that GLM-5.3, an open-weight AI model from Zhipu AI, possesses advanced cyber exploit capabilities comparable to safeguarded US models but lacks effective safeguards, increasing risks for misuse.

Naive-N0.5-Flash is an open-weight 309B Mixture-of-Experts AI model with 15.5B active parameters, optimized for coding and AI research, featuring a native 1M-token context window and high inference speeds.

Deem 9b is an open-weight AI model for calibrated decisions, released by LibertAI Labs as a competitor to TypeSafe's Jev, with honest benchmarks and open resources.

Xiaomi released open-weight MiMo V2.6 Pro and Flash models. The author switched their personal agent from DeepSeek V4.1 Flash to MiMo V2.6 Pro, finding lower costs and comparable or better results in testing.

OpenJudgement-4B-Preview is an experimental Qwen-based AI model fine-tuned for classification and scoring tasks, offering an open-weight alternative in early preview.

NVIDIA releases Nemotron 3 Diarization, an open-weight 100M-parameter model that achieves state-of-the-art speaker diarization with a 14.72% error rate, supporting real-time and offline processing for up to eight speakers.

A curated list of key AI founders, researchers, and builders for 2026, highlighting leaders from major AI labs and notable figures in the industry.

Jev-Omni is the first open-weight model to extend typed decisions to multimodal, supporting text, images, audio, and video, and directly returning the probability distribution of options without generating explanations.

This paper analyzes how open-weight AI demand affects the economic value of older NVIDIA GPUs, showing that open-weight models are cheaper and that hardware like the A100 remains cost-effective for specific workloads.

dots3-note Preview is a multimodal Mixture-of-Experts AI model with 280B total parameters and 16B activated, supporting up to 512K token context, released as an open-weight model on Hugging Face.

The article analyzes the shift in leadership for open-weight AI models from the U.S. to China, highlighting that Chinese models have surpassed American ones in commercial viability and download numbers.

CAISI's assessment finds that Z.ai's GLM-5.3 is the most cyber-capable open-weight AI model to date, but it still trails U.S. frontier models by approximately four months in capability.

The author shares a DIY Jev-like inference setup using open weight LLMs, demonstrating that simple prompting with logit-based verification achieves good accuracy without fine-tuning, and provides a rust web server for local deployment.

Jev-Omni is a multimodal decision classifier built on Gemma 4 12B IT, fine-tuned to handle text, images, audio, and video, providing probability scores for options based on questions.