Techyon

Your Agent Cluster Is Not a GPU Farm: The Hidden CPU Bottleneck in Agentic AI

The article argues that deploying expensive GPU clusters for autonomous coding agents often overlooks a hidden CPU bottleneck that limits performance. It states that the CPU, rather than the GPU, becomes the primary cons…

→ View original source

What's your favorite medium to large (~50-150b params) MOE model? Aside from Qwen flash next.

A Reddit user asked for recommendations on favorite medium-to-large (~50-150B parameter) mixture-of-experts models, aside from Qwen Flash Next, after upgrading RAM to run larger MOEs split across RAM and VRAM. They liste…

→ View original source

[New Model] Sage-A1.5 released based on Qwen3.6-35B MoE with marketed benchmarks: "highly competitive performance against frontier-scale systems such as Kimi-K3, GLM-5.3, and GPT-5.6-sol"

Sage‑A1.5 is a new language model built on the Qwen3.6‑35B MoE architecture and quantized with q4_k_m. The release claims highly competitive performance against frontier models such as Kimi‑K3, GLM‑5.3, and GPT‑5.6‑sol. …

→ View original source

25 reproducible audits of the metrics and gold labels behind LLM benchmark scores

Munim, Inc. developed "eval-reliability-audits," a catalog of 25 reproducible audits designed to identify defects in LLM benchmark scoring code and gold labels. Using sealed containers and pre-registered synthetic fixtur…

→ View original source

Kitten-TTS 2: New Open-source SOTA 1.58 bit TTS (CPU-only with voice cloning, multilingual, emotion control)

Kitten-TTS 2 is an open-source state-of-the-art TTS model that introduces the world's first 1.58-bit ternary weights for ultra-efficient speech synthesis. Running entirely on CPU with zero-shot voice cloning and multilin…

→ View original source

2-bit copy in VRAM + exact residual in system RAM as the verifier: speculative decoding for models that don't fit your GPU

This Reddit post introduces a speculative decoding technique for large language models that exceed GPU memory capacity. The method utilizes a 2-bit quantized copy in VRAM combined with an exact residual in system RAM to …

→ View original source

Agent S4: an open model beats GPT-5.5 on computer use by evolving its own tools ($1.10/task, fully open source)

Agent S4, an open-source agent scientist, enables a 27B parameter model to outperform GPT-5.5 and Opus 4.8 on computer-use tasks by autonomously evolving its own tool interfaces. Unlike fixed-interface agents that rely o…

→ View original source

OpenAI’s Rogue Agents Accessed Crypto Wallets As Anthropic Reveals $80 Billion Loss

OpenAI reportedly suffered a security breach involving rogue agents accessing cryptocurrency wallets, while Anthropic disclosed an $80 billion financial loss. These concurrent revelations underscore critical security vul…

→ View original source

models: Add Flint model (FlintForCausalLM) by alexey-rsi · Pull Request #30928 · ggml-org/llama.cpp

The Flint model (FlintForCausalLM) has been merged into llama.cpp via PR #30928, enabling local inference of this 39.8B parameter decoder-only mixture-of-experts transformer. It features 49 layers — 2 dense followed by 4…

→ View original source

The $15 billion ad-powered AI startup your doctor may already be using

OpenEvidence, a Miami-based AI-powered medical search platform founded in 2022 by Daniel Nadler, has rapidly gained traction among U.S. physicians, with nearly two-thirds of doctors using its chatbot to process over 42 m…

→ View original source

Show HN: Omnara – Open-source alternative to Claude Managed Agents

Omnara is an open-source managed agent platform (Apache 2.0 license) enabling self-hosted AI agents compatible with Claude Managed Agents and OpenAI’s Agents API. It supports any model via OpenAI/Anthropic-compatible end…

→ View original source

Using thousands of AI agents to uncover new leads in biological data

OpenAI has deployed thousands of AI agents to analyze complex biological data, aiming to accelerate the discovery of new research leads in the life sciences. The approach leverages large-scale agentic AI systems to proce…

→ View original source