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.
Step 5 Preview delivers competitive performance in agentic tasks and financial analysis with a 600B parameter MoE architecture, achieving a score of 44 on the Artificial Analysis Intelligence Index.
An 'Infinite-Parameter LLM' architecture has been proposed, which converts real-time interaction data into weights via a hypernetwork and performs online updates using a Bayesian approach.
China Telecom has released Xing4.0-29B-A4B, a 29B MoE model trained on Ascend NPU.
NVIDIA achieved 1,025 TFLOPs per GPU on GB200 for Dropless MoE training via Transformer Engine for JAX, improving throughput by approximately 10x.
Ant Group has released the finance-specialized open-weight model 'Ling-3.0-flash-Fin', achieving high efficiency with 5.1B active parameters.
Edge0 released an inference framework that runs 35B MoE models on 3GiB of memory using SSD expert streaming technology.
DeepSeek released the 552B MoE model V4.1-Flash, replacing V4-Pro and optimizing KV cache to 890 bytes per token.
Cohere has released 'North Small Translate', an open-weight MoE-based translation model that outperforms proprietary models on the WMT26 benchmark.
Sber released the MoE-based GigaChat-3.5-Reasoning, achieving DeepSeek V4 Flash-level performance while reducing inference token usage by 37%.
DeepSeek has released V4.1-Flash, which offers superior performance and efficiency compared to V4-Pro, along with reduced API prices.
inclusionAI has released Ling-3.0-flash-VL, a multimodal model featuring a 1M token context window and MoE architecture.
Cohere released a Megakernel serving engine for the North Mini Code model, achieving up to 1.41x end-to-end inference speedup over vLLM on a single H100 in BF16.
IFM has open-sourced K2 Horizon, a suite of 6 models that exposes the entire process from pre-training to agent post-processing.
Vercel AI Gateway is offering the medical-specific model Ling 3.0 Flash Sante for free until October 4.
llama.cpp now supports running the Nemotron-3-Puzzle-75B-A9B model, which adopts NVIDIA's hybrid MoE architecture.
The Apple Silicon-specific inference engine Lily achieved a 35% improvement in decoding speed and high hardware utilization compared to MLX-LM.
Tencent has released Hy4 Preview, an open-weight LLM with 770B parameters and 1M context support.
NVIDIA has released the NVFP4 quantized version of DeepSeek-V4-Pro-0813, supporting a MoE architecture that activates 400B parameters out of a total of 1.65T.
The GLM-5.3 model has been visualized on HF Viewer, confirming that while its architecture remains identical to GLM-5.2, performance has improved through training enhancements.
Qwen3.8-Flash-Next demonstrates superior performance compared to existing models at 1/9 the cost.
LayerStoRm, an open-source engine that runs large MoE LLMs on consumer GPUs using PCIe streaming, has been released.
FreeToken proposes a system for efficiently serving frontier-grade MoE models on hardware comparable to gaming PCs.
FreeToken has released an edge-native serving engine that enables interactive-speed execution of frontier MoE models on consumer hardware.
AntLing has open-sourced six Ling-3.0-based models applying WSM technology.