Insider Brief Sygaldry Technologies, Inc. announced today that it has raised $139M in Series A and Seed financing to build quantum-accelerated AI servers. The $105M Series A in March 2026 was led by Breakthrough Energy Ventures, following a $34M seed round led by Initialized Capital. Sygaldry’s servers aim to exponentially speed up critical AI algorithms, […]
Nous Research has published Lighthouse Attention, a selection-based hierarchical attention mechanism that wraps around standard scaled dot-product attention during pretraining and is removed afterward. Unlike prior methods such as NSA and HISA that pool only keys and values, Lighthouse pools Q, K, and V symmetrically across a multi-resolution pyramid, reducing the attention call from O(N·S·d) to O(S²·d) and running stock FlashAttention on a small dense sub-sequence. Tested on a 530M Llama-3-style model at 98K context, it achieves a 1.40–1.69× end-to-end wall-clock speedup against a cuDNN SDPA baseline with matching or lower final training loss.
The post Nous Research Proposes Lighthouse Attention: A Training-Only Selection-Based Hierarchical Attention That Delivers 1.4–1.7× Pretraining Speedup at Long Context appeared first on MarkTechPost.
Running AI agents in a local script is straightforward. Running them reliably in production across teams, across restarts, with isolated environments per context is a different problem entirely. BerriAI, the company behind the LiteLLM AI Gateway, is now open-sourcing a purpose-built answer to that problem: the LiteLLM Agent Platform. The platform is described as a […]
The post Meet LiteLLM Agent Platform: A Kubernetes-Based, Self-Hosted Infrastructure Layer for Isolated Agent Sandboxes and Persistent Session Management in Production appeared first on MarkTechPost.
Zyphra’s latest release shows that an autoregressive MoE model can be converted into a discrete diffusion model with no systematic loss in evaluation performance. ZAYA1-8B-Diffusion-Preview achieves up to 7.7x inference speedup over autoregression by shifting decoding from memory-bandwidth bound to compute-bound — a key advantage as modern GPUs continue scaling FLOPs faster than memory bandwidth.
The post Zyphra Releases ZAYA1-8B-Diffusion-Preview: The First MoE Diffusion Model Converted From an Autoregressive LLM With Up to 7.7x Speedup appeared first on MarkTechPost.
Speculative decoding was already clever. DFlash makes the draft stage parallel, turning diffusion from a clumsy text generator into a very…Continue reading on Medium »
The MIPI Alliance, an international organization that develops specifications that standardize wired interfaces for mobile and other connected ecosystems, has announced the formation of a “Physical AI Birds of a Feather (BoF) group” dedicated to explor…
Hay un momento muy específico en cualquier producto de IA en tiempo real que no aparece en ningún benchmark. No está en los dashboards, no…Continue reading on Medium »
Nous Research releases Token Superposition Training (TST), a two-phase pre-training method that cuts wall-clock training time by up to 2.5x at matched FLOPs by averaging contiguous token embeddings into bags during Phase 1 and reverting to standard next-token prediction in Phase 2 — without changing the model architecture, tokenizer, optimizer, or inference-time behavior. Validated at 270M, 600M, 3B dense, and 10B-A1B MoE scales.
The post Nous Research Releases Token Superposition Training to Speed Up LLM Pre-Training by Up to 2.5x Across 270M to 10B Parameter Models appeared first on MarkTechPost.
63% of organizations have no AI governance policy. Shadow AI is already running inside your stack — here is the data.
The post Enterprise AI Governance in 2026: Why the Tools Employees Use Are Ahead of the Policies That Cover Them appeared first on MarkTechPost.
Fastino Labs has released GLiGuard, a 300M parameter open-source safety moderation model that evaluates four safety tasks — prompt safety, jailbreak strategy detection, harm category classification, and refusal detection — in a single forward pass. Built on an encoder architecture rather than the decoder-only design used by most guardrail models, GLiGuard achieves up to 16x higher throughput and 16.6x lower latency than current state-of-the-art models, while matching or exceeding the accuracy of models 23 to 90 times its size across nine safety benchmarks. Model weights are available under the Apache 2.0 license on Hugging Face.
The post Fastino Labs Open-Sources GLiGuard: A 300M Parameter Safety Moderation Model That Matches or Exceeds Accuracy of Models 23–90x Its Size appeared first on MarkTechPost.