NVIDIA Newsroom published this on Nov 8, 2026 (source: nvidianews.nvidia.com).
NVIDIA's board announced an independent investigation into recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort. A committee of independent directors assisted by outside independent counsel will oversee the inquiry. The board said it has no reason to believe national security or NVIDIA secrets were compromised, and that Huang fully supports the review and has committed to cooperating. The statement follows reports that China recruited an AI executive in an attempt to spy on Huang.
OfficialOpenAI and ModRetro launch a Codex-built Game Boy Color game collection for the Chromatic
OpenAI Developers published this on Nov 8, 2026 (source: developers.openai.com).
OpenAI and ModRetro are collaborating on game creation for the ModRetro Chromatic handheld: a collection of 13 games built with Codex for Game Boy Color, playable in the browser or downloadable as ROMs for compatible hardware. The Codex-branded Chromatic is slated to reach OpenAI's Supply Co. store on November 9, per the announcement.
OfficialArtificial Analysis ships Decision Index v2 with 45% private datasets and calibration-aware scoring
Artificial Analysis published this on Nov 8, 2026 (source: artificialanalysis.ai).
Artificial Analysis released Decision Index v2, its second generation of the proprietary decision-model evaluation: the private-dataset share rises from 25% to 45%, and a new Decision Score rewards both correctness and calibrated confidence (right answers score above 50, wrong answers below 50, with confidence moving the score within each half). Drex Max v1.1 from Nace AI narrowly leads decision models at 73, with GLiDE v1.2 (Fastino Labs) and Mammoth 27B (Baseten) within a point.
OfficialNous Research ships hermes-security, an official security-review plugin for Hermes Agent
Hermes Marketplace published this on Nov 9, 2026 (source: hermes-agent.nousresearch.com).
Nous Research published hermes-security in the official Hermes plugin catalog: the agent builds a threat model of a local repository, maps attacker-reachable paths and trusted inputs, splits code into chunks reviewed by parallel subagents, and traces whether an attacker can reach each candidate finding. Reports keep evidence, rejected candidates and coverage gaps, and export to Markdown, JSON, CSV and SARIF. Optional reproduction tests run in a temporary copy of the repository, and the plugin requires Hermes 0.21.5 or later.
OfficialNVIDIA broadens its Korea push: Samsung physical AI, Hyundai agent safety, and Dynamo recipes for Korean open models
NVIDIA Blog published this on Nov 8, 2026 (source: blogs.nvidia.com).
NVIDIA published three updates from its Korea ecosystem push. Samsung Electronics is expanding semiconductor collaboration with NVIDIA on accelerated computing and AI physics spanning materials research through manufacturing. Hyundai Motor Group and AIM Intelligence are exploring NVIDIA NemoClaw blueprints and the OpenShell runtime for secure enterprise agents, part of the NVIDIA Open Agent Safety Platform, while LG AI Research opened an API for K-EXAONE-2.5 trained on Nemotron datasets. NVIDIA also released Dynamo-optimized deployment recipes for Korea's leading open models - SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, Upstage's Solar Open 2 and Motif Technologies' Motif-3 - for multi-node inference at data-center scale.
OfficialNVIDIA, Rebellions and SK Telecom announce sovereign AI infrastructure collaboration
NVIDIA Newsroom published this on Nov 8, 2026 (source: nvidianews.nvidia.com).
Rebellions plans to connect its R200 NPUs through NVIDIA NVLink Fusion and use MGX rack infrastructure, while SK Telecom plans to deploy the systems in its AI data centers. The collaboration is supported by Korea’s Ministry of Science and ICT. R200 systems are expected to be ready for deployment in 2028; this is a collaboration announcement, not a delivered system.
OfficialBaidu introduces ERNIE Evolve for automated AI research and inference optimization
Baidu ERNIE Research published this on Nov 9, 2026 (source: ernie.baidu.com).
Baidu describes ERNIE Evolve as an automated research system that selects research directions, edits code and runs experiments in isolated environments. Separate evaluator and reviewer components decide which candidates the system retains. Baidu reports applying it to public training and kernel benchmarks and to inference optimization for the in-development ERNIE-5.5-Flash; the reported results are from Baidu, not independent testing.
Officialllama.cpp pre-release adds CPU sharing and an automatic tuner for MoE cache misses
Llama.cpp releases published this on Nov 7, 2026 (source: github.com).
The llama.cpp b12191 pre-release adds --moe-cache-cpu-share to compute a selected share of cache misses on the CPU instead of uploading all missing experts to the device. A new tuning script runs a server at different cache sizes and CPU shares to recommend settings for the local machine. This is an upstream release-note description, not an independently measured speedup.
Officialllama.cpp pre-release fixes tied radix Top-K selections on CUDA and Metal
Llama.cpp releases published this on Nov 8, 2026 (source: github.com).
The llama.cpp b12208 pre-release updates radix Top-K handling on CUDA and Metal to make tied selections deterministic rather than dependent on thread scheduling. Its Metal implementation replaces atomic compaction with deterministic ranking and coalesced accesses. This concerns the radix Top-K selection path, not a guarantee that entire model outputs are deterministic. Any performance figures in the release notes are upstream measurements, not independently verified benchmarks.
OfficialInkling 2 Small becomes free on OpenCode for a limited time
@soumithchintala on X published this on Nov 7, 2026 (source: x.com).
Thinking Machines’ Soumith Chintala announced limited-time free access to Inkling 2 Small through OpenCode. OpenCode lists the free model but says its provider retains submitted data and may use it to monitor performance and train models. The free offering is unavailable in the European Union and other listed regions. This is an access offer, not a new model release; no end date was stated.
OfficialTencent publishes DSpine, a speculative-decoding draft model for Qwen3.8-27B
Hugging Face @tencent published this on Nov 7, 2026 (source: huggingface.co).
Tencent posted Qwen3.8-27B-DSpine-b16 on Hugging Face, a five-layer DSpine draft model for Qwen3.8-27B that generates draft tokens in parallel by passing predicted predecessor features to adjacent positions. The checkpoint uses block size 16 and an 8,192-token sliding window, is tagged for SGLang speculative decoding, and is released under Apache 2.0. Tencent’s card reports longer acceptance lengths than official DFlash2 on GSM8K, MATH-500, HumanEval, MBPP, and MT-Bench, and on WorkBuddy agent suites, measured with SGLang 0.5.22.
OfficialNVIDIA announces planned investments in Korean startups Point2, Motif, RLWRLD, and Upstage
NVIDIA Blog published this on Nov 9, 2026 (source: blogs.nvidia.com).
NVIDIA and Korea’s Ministry of Science and ICT hosted the MSIT-NVIDIA AI Startup Partnership Forum at COEX, where Vice Minister Ryu Je-myung and NVIDIA leaders met more than 20 Korean startups across foundation models, physical AI, AI infrastructure, and AI semiconductors. NVIDIA said it is investing in interconnect startup Point2 Technology and that additional investments are planned in Motif Technologies, RLWRLD, and Upstage. Point2 builds radio-frequency chip-to-rack interconnects as an alternative to copper and optical links; Motif develops foundation models including the open-weight Motif-3; RLWRLD works on robotics foundation models for dexterous manipulation; and Upstage builds Solar language models and document AI tools. NVIDIA tied the planned investments to more than 1,000 Korea-based members of its Inception program.
Official by 36KrAlibaba Cloud raises Token Plan personal quotas by about 70%+ without changing Lite/Essential/Standard monthly prices
36Kr published this on Nov 9, 2026 (source: 36kr.com).
Alibaba Cloud said personal-edition Token Plan monthly allowances rose about 74%–78% for Lite, Essential and Standard while those monthly prices stayed the same; Pro quota and price were also adjusted. Primary notice is Alibaba Cloud’s WeChat post; 36Kr is secondary.
Official by 36KrQwen Office 2.0.0 adds Team collaboration so multiple people can work with an agent in one shared space
36Kr published this on Nov 9, 2026 (source: 36kr.com).
Qwen Office client release notes for 2026-11-09 list version 2.0.0 with Collaboration (Team Agent), human-AI co-creation, materials workspace, automated/triggered tasks, and three-client connectivity. 36Kr is secondary coverage of the same ship.
OfficialByteDance Seed releases Seedream 5.5 with stabler multi-round edits, native RGBA, and clearer text
ByteDance Seed published this on Nov 9, 2026 (source: seed.bytedance.com).
ByteDance’s Seed team says Seedream 5.5 improves multi-round editing with pixel-level preservation outside edit regions (up to 100 rounds), adds native RGBA transparent asset generation, and lifts photorealism and small-text rendering versus Seedream 5.0 Pro. The model is rolling out on Jimeng, Doubao, Doubao Work, and Volcano Ark; a later Pro variant in the 5.5 series is previewed.
