The Hallway Track
11 tracked signals on mixture-of-experts.
Moonshot AI's Kimi K3 is a 2.8T-parameter open-weight model matching frontier closed models on coding benchmarks.
“it has OpenAI and Anthropic terrified because its Trust Me Bro benchmark performance is on par with and in some cases beating Claude Fable and GPT 5.6 Soul”
Zhipu's GLM 5.3 Flash anonymously dominated Open Router and costs 40x less than Claude
“at its peak, OX Alpha accounted for nearly a third of Open Router's entire weekly traffic”
Thinking Machines releases Inkling, a 970B-parameter Apache-licensed open-weights multimodal model.
“the weights are Apache licensed and sitting on hugging face right now”
Reflection AI launches Beam, a 501B/23B-active US-trained open MoE model for coding and agentic work
Alibaba previews Qwen4 architecture via 176B MoE model with 1M token context
Z.ai's 753B-parameter GLM 5.3 MoE model is now available on Amazon Bedrock for enterprise use
NVIDIA details Mixture-of-experts training methods to efficiently scale biological foundation models.
“Mixture-of-experts (MoE) architectures take a different approach to scaling by using many subnetworks, or experts, while activating only a small subset for each token.”
Qwen 3.8 Flash Next, a free open-weights MoE model, rivals paid closed AI systems.
“it seems that we can switch out paid closed systems to open weights AI that we can download and run ourselves forever”
Google DeepMind's open-weight Gemma 4 model family is now available on Amazon Bedrock.
“Artificial Analysis reports an Intelligence Index of 39 for Gemma 4 31B, well above the median of 15 in the 4B–40B open-weights class.”
NVIDIA details advanced fusion kernels that boost training throughput for mixture-of-experts (MoE) models.
“Mixture-of-experts (MoE) models have quickly become a foundational component of modern, large-scale AI systems.”
JetBrains released Mellum2, a 12B mixture-of-experts model, on Hugging Face.