Proposes structure-preserving pruning for mixture-of-experts models, introducing PEEP and DCEP to keep important structure while reducing active experts; shows empirical gains at 50–75% retention on multiple benchmarks.
Jev Router chooses a model and reasoning effort for each request based on task needs, cost policy, and model preferences. The documentation explains its optional plugin, cost tiers, model restrictions, supported request formats, and reasoning-effort settings.
A Shenzhen AI startup, Hypersphere Technology, secures a large angel round from several investors led by Legend Capital; co-founded by a CUHK professor, it pursues “Infinite Models” that coordinate specialized models with both in-house and external components, using routing and aggregation to handle tasks and researching advanced training and merging techniques.
Ling-3.0-flash-Sante is a healthcare-focused research model with 124B parameters (5.1B active per token), built on Ling-3.0-flash. It supports long-context analysis and tool-assisted workflows, but outputs are not medical advice and require expert review and regulatory compliance. It combines a hybrid MoE architecture for efficiency and is evaluated on professional healthcare tasks and safety benchmarks.
The document announces HuggingFace Transformers release v5.20.1, introducing a new model Beam from Reflection AI optimized for coding and agentic tasks, designed to run on user hardware or cloud. It also lists documentation updates and multiple bug fixes related to torch compatibility, model loading, and MoE/parallelism improvements.
DeepSeek V4.1 Flash 1017 is a cost-efficient sparse MoE model with image understanding, fast throughput, and KV caching, priced at $0.15 or $0.60 per 1M tokens; the item is featured by LAXIMA AI Signal from Hacker News with community engagement and a link to the original source.
The paper introduces DepthGraft, a hierarchical KV-sharing design intended to improve information flow in deep language models while reducing KV cache size by about one third. Tests on models from 0.5B to 16B parameters report downstream improvements, including gains on MMLU-Pro, CMMLU, and C3 for a 16B MoE model.
The document reviews various model-routing approaches to reduce inference costs by assigning tasks to the most suitable model, considering cost, quality, and latency, and cites multiple vendor implementations and strategies.
LatentLoop releases pretrained, multi-scale MoE language model checkpoints that use asymmetric compute allocation to expand model width while compressing sequence length on the first pass. It manages separate KV states per visit, provides token-budgeted evaluation, and offers PyTorch-based loading and inference with optional FlashAttention backends.
The paper proposes a Mixture-of-Experts approach for multi-agent pathfinding, training a router to selectively combine specialized experts for different map distributions. Results show MoE often improves over a pretrained policy and outperforms current methods, though well-tuned monolithic models can still be superior in some scenarios.