mixture-of-experts news, stories and analysis

Donut chart showing Qwen3.8-Max mixture-of-experts parameter distribution. Only 95 billion parameters are active per request out of 2.4 trillion total, meaning 3.96 percent of the model computes for any single token.

Alibaba Qwen3.8-Max narrows open-weight frontier gap

Qwen3.8-Max is Alibaba's 2.4T open-weight model ranking fifth in text and second in vision on Arena. The gap with US frontier labs narrows but does not close.

Lars Cornelissen · Aug 3, 2026

Benchmark performance comparison showing Kimi K3 scoring 59.2 on SWE-bench Verified against Claude Fable 5 at 70.4 and GPT-5.6 Sol at 72.1, with Kimi K2 at 51.8. Kimi K3 reaches 83 percent of the top proprietary score.

AI

Kimi K3 hits open frontier with 2.8T mixture of experts

Kimi K3 is an open-weight 2.8-trillion-parameter mixture-of-experts model from Moonshot AI. It ranks fourth of 580 models on Artificial Analysis, behind only three proprietary models, while activating just 104B parameters per token.

Lars Cornelissen · Jul 31, 2026

Open-weight model comparison showing Inkling at 975B total parameters and 41B active, Inkling-Small at 276B total and 12B active, and Hy3 295B at 295B total and 21B active

AI for dummies

Inkling 975B: Thinking Machines releases open-weights model

Inkling is Thinking Machines Lab's first open-weights model. At 975B parameters with 41B active and Apache 2.0 licensing, it targets fine-tuning, not frontier benchmarks.

Lars Cornelissen · Jul 17, 2026

Abstract data visualization representing LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion total parameters and 48 billion active parameters per token, released under MIT license on July 5 2026.

AI for dummies

LongCat-2.0 open weights: what 1.6 trillion parameters means

LongCat-2.0 is a 1.6 trillion parameter AI model from Meituan that activates only 48 billion parameters per token. Its weights are now open under the MIT license.

Lars Cornelissen · Jul 5, 2026