AI Briefing#moe
The latest AI and developer news about #moe, with the original source and a short summary.
The latest AI and developer news about #moe, with the original source and a short summary.
Precision Numerics releases Boole-32B-A2.5B-Preview, a 1.8-bit MoE model that is 9x smaller than BF16 and 2x faster than INT4.
InfraRSI achieved a geometric mean speedup of 1.83x on H100/H200 by simultaneously exploring configurations and code.
The guide maps Claude Code's default model IDs to the 560B-total, 25B-active multimodal MoE via an Anthropic-compatible interface.
Built on llama.cpp, the tool automatically allocates model weights to VRAM, RAM, or flash storage to maximize prefill and decode speeds on heterogeneous hardware.
The 560B-parameter model supports text, image, and video inputs and achieves 92.21 on OmniDocBench 1.5.
Inkling 2 Small achieved the final score of the first generation in just 75 steps of RL fine-tuning.
The 45B parameter model processes audio, video, and text in 160ms chunks while delegating complex reasoning to a larger backend model to maintain low latency.
The model achieves 75.00 on SWE-bench Verified and 90.00 on AIME2026, supporting 256K context natively.
Mistral released the weights for its 1T MoE model, Mistral Large 4, demonstrating token efficiency.
Sleep improved the AIME-24 score by 5.4%p through offline integration.
The optimization achieves 5.7% to 6.1% gains on RTX 5090 for Qwen3 and Qwen3.6 models.
The 1.58-bit MoE model achieves 172 tok/s decode and 7000+ tok/s prefill speeds in both Stock llama.cpp and the PrismML fork.
StepFun released the weights for its 600B MoE model Step 5 Preview and confirmed it can run on 8 H200s.
Apple's 20B 'Core Advanced' model runs on macOS 27 using approximately 2.0–3.4 GiB of RAM by storing weights on flash, consuming about one-third of the energy of comparable open models.
PyTorch 2.15 improves Mac GPU MoE decoding by 5-6.5x and supports FP8.
The system uses Lookahead Expert Scheduling to preload weights, achieving 6.07x speedup over the baseline with only a 0.16-point benchmark drop.
The Raon-1 mini preview, a 30B-class MoE model, has been released under the Apache 2.0 license.
The 180B parameter hybrid attention MoE model features a 1,048,576-token context window and selectable reasoning effort levels.
The model activates 6.5B parameters per inference step and supports speech recognition, audio reasoning, and synthesized speech responses under an MIT license.
Trained on 16,000 GB200 GPUs for 38 days with 50T tokens, claiming 4–12x compute efficiency improvement over DeepSeek-V4-Pro-Base
Kolibri-1 is a 78B-parameter Mixture-of-Experts model with 3.46B active parameters, optimized for German and English reasoning.
The update enables Multi-Token Prediction for the 12B-parameter Mellum2.1 Thinking model, which features 2.5B active parameters and a 131k context window.
The update delivers up to 138% token generation per watt improvements for MoE models and introduces experimental hierarchical KV cache offloading.
The Apache 2.0-licensed model features 23B active parameters, a 1M token context window, and day-0 support for vLLM, SGLang, and Transformers.