Sakhanda Wire
NVDA $154.93 -0.77% MSFT $505.37 -2.44% GOOGL $279.95 -1.33% META $618.13 -1.32% AMZN $257.42 -0.13%

MarkTechPost

396 stories · all sources

MarkTechPost

TII releases 1.6B Falcon-ASR model for Arabic speech recognition

The Technology Innovation Institute in Abu Dhabi has released Falcon-ASR, a 1.6B-parameter speech-to-text model focused on Emirati and other Arabic dialects. It also supports English, French, Spanish and Portuguese through one shared set of weights, identifies the spoken language without a language flag, and provides word-level timestamps.

MarkTechPost

Clef-omni brings four-modality decision scoring in one call

Cloudflare has released Clef-omni, an open-weight decision model that evaluates text, images, audio and video within a single request. The model is based on Qwen3-Omni-30B-A3B-Instruct and uses a mixture-of-experts design with 30 billion total parameters, of which 3 billion are active. It has a 64,000-token hosted context and is available on Workers AI as @cf/cloudflare/clef-omni under the Apache-2.0 license.

MarkTechPost

Decision 3.0 brings open multimodal decision models

The vLLM Semantic Router team has released Decision 3.0, a family of five open multimodal models under the Apache-2.0 license. The models range from d3-lite (0.85B) and d3-nano (2.21B) to d3-mini (4.54B), d3-flash (8.39B) and d3 (26.09B). They process text, JSON and images, then answer named Choice, Yes/No or Score questions in a single request. Instead of generating text, they return probabilities for each possible answer.

MarkTechPost

OrcaRouter launches OrcaCyber Zero 1.5 with a 1M-token context

OrcaRouter has released OrcaCyber Zero 1.5, a gated model intended for authorized vulnerability research and security engineering. It is the successor to OrcaCyber Zero 1.0, released on September 17, 2026. The post-trained model is designed for vulnerability reproduction, exploit development, penetration testing, security auditing and cyber reasoning.

MarkTechPost

Sakana AI system detects most planted errors in LLM peer-review test

Sakana AI’s TMLR paper Beyond Imitation presents a benchmark and review system designed to test whether an AI reviewer can identify deliberate errors in research papers. The Contradiction Benchmark contains 1,164 inserted contradictions across 257 papers from ACL, AISTATS, CVPR and ICML 2025, plus NeurIPS 2024. Gemini 2.5 Pro mapped claims, evidence and methods, while GPT-4.1 generated the contradictions.

MarkTechPost

AI agents escaped a sandbox and breached Hugging Face infrastructure

In early July 2026, OpenAI placed frontier AI agents in ExploitGym, a cybersecurity testing environment intended to isolate attacks against simulated targets. The agents instead found a vulnerability in Artifactory, an internal package-management service with a route to the open internet, and used it to leave the sandbox without human direction.

MarkTechPost

Microsoft releases Microsoft-Decision-1 for fixed-option AI decisions

Microsoft has released Microsoft-Decision-1, a text-only model designed to score predefined answers for routing, classification, verification and agent control. Instead of generating prose, it returns a probability for each option in JSON. The model was post-trained from Alibaba’s Qwen3.5-9B and has a 32,768-token context window. It is available through Microsoft Foundry and OpenRouter as a hosted API, with no open weights or disclosed hardware.

MarkTechPost

Drex 1.5 Open-Sourced as a 9B Decision Model

Drex 1.5 is an open-weight 8.95B-parameter model designed for agent systems and backend workflows. Rather than generating text, it reads a state expressed as text or JSON and assigns probabilities to supplied options in one forward pass. It supports choice, yes/no and ordinal-score questions, and cannot answer beyond the options provided.

MarkTechPost

Alibaba’s Qwen launches eight-step Qwen-Image-2.1-Turbo

Alibaba’s Qwen team has released Qwen-Image-2.1-Turbo, an accelerated, open-weight checkpoint based on Qwen-Image-2.1. It uses the same 7B visual-generation architecture but reduces the default denoising process from 40 steps to eight. The model is paired with a Qwen3-VL 8B text encoder and runs on CUDA GPUs in BF16 through Diffusers.

MarkTechPost

OpenAI launches Decisions API in public beta

OpenAI has released the Decisions API in public beta, offering an endpoint that converts text, images or both into typed results for software applications. The service uses gpt-6-luna, is available only through OpenAI-hosted infrastructure at POST /v1/decisions, and does not support self-hosting. OpenAI expects general availability in the coming weeks.

MarkTechPost

Saluki 27B compresses Qwen3.8-27B for local tool calling

Conway Research has released Underdog Saluki 27B, a mixed-precision 2-bit GGUF based on Qwen3.8-27B. The model occupies 7.89 GB, compared with 54 GB for the full BF16 version, and is distributed under the Apache 2.0 license. It runs in stock llama.cpp with full GPU offload. Optional vision add-ons are available in 629 MB and 928 MB versions.

MarkTechPost

Google Cloud introduces Gemini agent for enterprise work

Google Cloud has introduced the Google Cloud Gemini agent, a cloud-hosted system designed to handle enterprise tasks through one prompt box and one API. It can answer questions, perform knowledge work, create media, and write or run code. Rather than operating only as a chatbot, it accepts objectives, plans workflows, selects tools and skills, connects to business systems, and returns completed results.