Signal
6 articles covering "mixture of experts"
Aleph Alpha published Kolibri on Hugging Face on October 3 under Apache 2.0; its own tech report puts the model on the quality-versus-serving-cost Pareto frontier.
StepFun's new Step 5 Preview outscores Google's Gemini 3.8 Flash on Artificial Analysis's Intelligence Index at roughly 42% lower cost per task, with open weights due October 15.
DeepSeek's new V4.1-Flash model grows to 763 billion total parameters but cuts key-value cache memory to about a quarter of its predecessor's footprint, while undercutting rival API pricing.
The 550-billion-parameter model activates 55 billion parameters per token, ships under the Linux Foundation's OpenMDW-1.1 license, and trades blows with China's Kimi-K2.6 on coding benchmarks.
Nvidia open-sources a 30B mixture-of-experts model that activates only 3B parameters yet matches DeepSeek's 671B model on olympiad-level math and coding benchmarks.
Mistral releases Small 4, a 119B-parameter open-weight model under Apache 2.0, and Forge, an enterprise platform for training custom models on proprietary data, as open-weight AI closes to within three months of frontier systems.