Tag: Mixture of experts

NVIDIA Cosmos 3: one world model for physical AI

At SIGGRAPH 2026 NVIDIA showed Cosmos 3 — a physical-AI foundation model fusing understanding, world generation and action. The edge Cosmos 3H (4B params) runs on Jetson, and Cosmos Dreams cuts simulation cost from 64 GPUs to one.

GLM-5.2 is 5× cheaper than Opus 4.8 — at a coding-score cost

Z.ai's Chinese GLM-5.2 model costs about five times less than Anthropic's Opus 4.8 for output tokens in the API, but clearly trails it on the hardest coding tasks. It shows where the price-versus-quality line runs in AI-assisted coding today.

Thinking Machines reveals Inkling — Mira Murati open-weight MoE 975B

Thinking Machines Lab, Mira Murati startup, released its first AI model — Inkling. An open-weight MoE with 975 billion total parameters, natively multimodal, trained on 45 trillion tokens. The company argues what matters is not the model itself but the ability to customize it.

LingBot-VA 2.0: robot control model trained on physics, not language

Robbyant released LingBot-VA 2.0, a robot control foundation model trained on physical causality from scratch. A new VAE tokenizer, multi-chunk prediction, and asynchronous inference pipeline deliver 93.6% on RoboTwin 2.0 and 225 Hz control frequency.

Z.ai launches ZCode: free agentic IDE powered by GLM-5.2

Z.ai (formerly Zhipu AI) launched ZCode — a free agentic development environment powered by GLM-5.2, trained entirely on Huawei chips. The tool costs up to 82% less than Claude Code and directly competes with Cursor, GitHub Copilot, and Google Antigravity.

Tencent opens Hy3 — a 295B MoE model under Apache 2.0

Tencent has released Hy3, an open Mixture-of-Experts model with 295B total and 21B active parameters, under Apache 2.0. In a blind expert evaluation it edged out GLM-5.1, sharpening the contest among Chinese open models.

Meituan open-sources LongCat-2.0: 1.6T MoE on Chinese ASICs beats GPT-5.5

Meituan unveiled LongCat-2.0, a 1.6T-parameter MoE trained without Nvidia GPUs, which spent two months anonymously leading OpenRouter as "Owl Alpha". It scored 59.5 on SWE-bench Pro, above GPT-5.5.

DiffusionGemma: Google DeepMind opens a new era of local text generation

Google DeepMind released DiffusionGemma — an open-source model that generates 256 tokens in parallel and self-corrects errors. On an NVIDIA H100 GPU, it achieves 4x higher speed than standard models, at the cost of lower generation quality.

Thinking Machines debuts: TML-Interaction-Small merges intelligence with real-time responsiveness

Mira Murati's startup has released its first foundational model — TML-Interaction-Small. The 276B-parameter MoE model is trained from scratch to handle audio, video, and text in 200-millisecond micro-turns, without any external dialog-management components.

MiniMax M2.7 — model językowy zdolny do udziału we własnym treningu

Mixture of Experts (MoE) — architektura, która pozwala budować coraz większe modele AI bez proporcjonalnego wzrostu kosztów obliczeniowych

GLM-5: duży model językowy zaprojektowany do autonomicznego inżynierowania oprogramowania

Google otwiera Gemma 4: Nowa generacja modeli AI stawia na lokalne działanie i agentowość