Search “NVIDIA”

SoMark open-sources the 1.3B document parsing model OXR-1.0

SoMark has open-sourced OXR-1.0, a 1.3B-parameter document parsing model that can convert PDFs and page images into Markdown or JSON. The weights have been released on Hugging Face, and lightweight testing can run on a single 24GB GPU. According to Rohan Paul, the model ranks first on olmOCR-Bench, OmniDocBench, and DocParsingBench, achieving a decoding speed of 558 tokens/s on a single NVIDIA RTX PRO 6000 GPU and parsing 100 pages end-to-end in about 101 seconds; its post-training stage employs a recursive self-improvement (RSI) loop, in which a general-purpose VLM judges parsing results and generates newly annotated pages, eliminating the need for manual annotation. QbitAI reports that the OXR main model open-sourced by the SoMark team has only 1.3B parameters and topped all three benchmark tests, with full parsing speed exceeding that of PaddleOCR-VL-1.6 and MinerU2.5-Pro, which were also evaluated in the same comparison. It covers 13 categories of document elements, including text, tables, formulas, chemical structure formulas, and code blocks, and can parse reading order and output layout coordinates.

Survival LLM releases offline AI emergency assistant

On October 28, 2026, Survival LLM released a free, open-source AI assistant and emergency reference library that runs entirely locally, and can still answer questions about food storage, burn treatment, water purification, and more after going offline, listing the articles its answers are based on. When the app starts, it detects the computer's memory and downloads a matching model, ranging from IBM Granite 4.2 8B on devices with 16–31 GB of memory to NVIDIA Nemotron 3.5 Super on Macs with 128 GB or more; the reference library is about 9 GB and contains roughly 930,000 Wikipedia articles, FEMA and CDC guidelines, military medical manuals, and iFixit repair guides. According to the report, in a blind test of 22 power-outage and first-aid questions, Nemotron 3.5 Super and Qwen3.8 27B each answered 21 correctly, while Granite 4.2 8B answered 14 correctly. There are currently no further developments.

Meta plans to raise up to $95 billion through an equity offering

On October 26, 2026, Meta Platforms announced plans to raise up to $95 billion through an equity offering, to be used for AI infrastructure, computing power, and personal superintelligence. The first round of transactions totaled $35 billion, of which about two-thirds came from common stock, and $12.5 billion in Series A depositary shares were issued; subsequently, it plans to raise up to an additional $60 billion through an at-the-market offering. NVIDIA, AMD, and Broadcom expressed interest in subscribing to a combined total of up to $5 billion in the Class A offering, but this is non-binding. On October 27, AMD announced a deep strategic partnership with Meta and invested $1 billion in Meta's equity offering. On October 28, Meta's $25 billion common stock offering was about five times oversubscribed.

Meta Shares AI Networking Innovations at OCP Summit

At the 2026 OCP Global Summit, Meta unveiled networking innovations supporting its latest AI models and its gigawatt-scale Prometheus cluster. The first liquid-cooled network rack introduced by Meta, IcePack, debuted alongside the 102.4T switch IceCube (128×800G, available with either Broadcom Tomahawk 6C or NVIDIA Spectrum-6). For scale-out, the Flattened Planar Fabric connects each GPU to eight network planes, simplifying the Clos architecture from 3 tiers to 2; for scale-across, Meta introduced BAG, a petabit-class super backbone connecting buildings, campuses and regions. Meta also introduced Tahan, an MTIA 300 Ethernet scale-up switch based on Broadcom Tomahawk5, and open-sourced its MetaRoCE software implementation, with the first user-space version, libsoftmetaroce, now available on GitHub.

Block releases full-duplex speech model Dessa

Block's internal AI research lab released the full-duplex speech-to-speech model Dessa on October 23. According to Unite.AI, the model was trained entirely from scratch on human conversational speech, can listen and speak simultaneously, and natively perceives emotion, sarcasm, and laughter; its inference is faster than real time and it can run locally on a MacBook Pro M4 Max or higher configuration, with the team completing training on a single NVIDIA GB200 NVL72. Block has released demo recordings and a public waitlist, and has opened early access to a small number of AI safety researchers. Other reports note that the model does not rely on a text intermediary and supports full-duplex conversation with the ability to interrupt and interject at any time, and that training consumes hundreds of times fewer tokens than large language models, but the author points out that this is currently only a small-scale research preview from Block's internal lab, the demos were selected by the company, and questions remain about how pure audio models perform on complex multi-step logical reasoning.

DeepSeek open-sources inference framework DeepInfer

On October 20, 2026, DeepSeek open-sourced its internal inference framework DeepInfer. According to Yuchen Jin, citing data disclosed by DeepSeek, a single cluster processes 155 trillion tokens per day and generates $38 in revenue per GPU-hour (off-peak pricing); DeepInfer supports both NVIDIA Blackwell and Huawei Ascend. A subsequent 36Kr report stated that its accompanying open-source deepseek-recipe supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. On October 21, swyx, citing Latent.Space data, said a single cluster processes 155T tokens per day with a 96% input cache hit rate, and argued that the lifetime value of each sold GPU has risen by about 30%. On October 24, Nathan Lambert said DeepSeek's single cluster processes 150T tokens per day, exceeding the combined open-source model inference volume of OpenRouter, Fireworks, Together, and BaseTen, and noted that the actual number of clusters is unknown; on the same day, LMSYS published a same-hardware benchmark comparison, stating that SGLang meets or exceeds DeepInfer on DeepInfer's own benchmark, with prefill throughput up to 1.04x and decode step length shortened by 1.1%–7.4% on 64x GB300.

Artificial Analysis publishes AA-AgentPerf results for Muse Spark Open

Artificial Analysis on October 24, 2026 published on AA-AgentPerf preliminary test results for Meta's newly open-weight model Muse Spark Open. The model is a 816B-parameter, 42B-active MoE, scoring 39 on the Artificial Analysis Intelligence Index. Tests show that NVIDIA B300 and GB300 each reach about 150M tokens/s per megawatt of power under throughput-first configurations; GB300 reaches about 108M and B300 about 92M at 100 tok/s output speed; AMD MI355X reaches 70M tokens/s per megawatt at 65 tok/s. Of the model's 62 layers, 46 use 2,048-token sliding window attention, reducing KV cache requirements at 1M context by about 75% versus full attention. Meta released an NVFP4 checkpoint, and AMD released an MXFP4 checkpoint. There are currently no subsequent development facts.

Schneider Electric launches prefabricated power modules for AI data centers

On October 23, 2026, Schneider Electric launched its next-generation prefabricated power modules and power skids for AI data centers. According to The Elec, the product is based on a reference design jointly developed with the Nvidia GB300 NVL72 AI server system, supports up to 142kW of power density per rack, and integrates power management and liquid cooling control. The product uses Galaxy VXL UPS technology and is offered in three capacities: 2MW, 2.25MW, and 2.5MW, with the 2.5MW version being its highest-capacity single-container prefabricated power solution to date. The product will be manufactured at a factory near Barcelona, Spain, launching first in Europe, entering North America later in 2026, and then expanding to Asia-Pacific. There are currently no subsequent developments.

ServeTheHome reviews the MSI XpertStation WS300

ServeTheHome reviewed the MSI XpertStation WS300. This machine is the second desktop workstation based on the NVIDIA DGX Station GB300, equipped with a Blackwell Ultra GPU with 252GB HBM3E (7.1TB/s), a 1600W power supply, and dual 400Gbps ConnectX-8 networking. The entire unit weighs nearly 30kg, uses a tower chassis with closed-loop liquid cooling, is priced above $100,000, and is positioned as server-grade hardware that can run local AI while sitting beside a desk. There are currently no subsequent developments.

heise tests RTX Pro 6000 performance loss at reduced TDP

heise online conducted an efficiency test of the Nvidia RTX Pro 6000 Blackwell, measuring the card's performance loss after reducing TDP. The test covered three power levels: 450 watts, 300 watts and 156 watts; the card has 96 GByte of VRAM and targets AI and workstation scenarios. The test used 450 watts, corresponding to the previous-generation GeForce RTX 4090, 300 watts, corresponding to the long-term power limits of earlier flagship cards, and 156 watts, corresponding to mid-range and lower-end cards such as the GeForce RTX 5060 or Radeon RX 9060 XT, as references. The report does not give the specific performance-loss figures for each level in the public summary; the full test data is available only to heise-Plus subscribers. This test is currently still the latest development, and there are no subsequent factual developments yet.

Doron Zauer proposes that the AI infrastructure layer is where the next batch of big winners will be found

On October 23, 2026, Doron Zauer, a partner at Earth & Beyond Ventures, offered his judgment that the next batch of big AI winners will come from companies building the infrastructure that giants rely on. He said that competition in the AI application layer is fierce and easily commoditized, while physical infrastructure such as silicon architecture, advanced materials, and connectivity technologies represents a long-term advantage that is hard to replicate. He noted that Israel has deep accumulated strength in chip design, photonics, and high-speed communications, and that entrepreneurs can target the architectural bottlenecks encountered by giants such as Nvidia, Apple, Intel, and Google. He also pointed out that commercialization success depends on yield, reliability, and system integration, and that early validation with industrial buyers and design partners is crucial. At present, this judgment is presented in a CTech report, and there are no other factual developments.

KT Cloud Breaks Ground on Ansan AI Data Center

On October 22, KT Cloud broke ground on its own 50MW AI data center in Ansan, Gyeonggi Province. The facility has a total floor area of approximately 36,878 square meters, is a nine-story above-ground, two-story underground building, and is scheduled for completion in 2029. According to The Elec, all server rooms in the data center will support liquid cooling and adopt a high-load structural design to accommodate Nvidia's next-generation Vera Rubin GPUs, while also introducing a natural cooling system to reduce cooling energy consumption. The project is part of KT Cloud's long-term strategy, which aims to achieve 1GW of AI data center capacity by 2031. There is currently no qualified factual evidence of subsequent progress.

Cohere announces public beta of Model Vault confidential computing

Cohere announced that the confidential computing feature of Model Vault has entered public beta. The feature is based on the NVIDIA H100 GPU confidential computing capability on Microsoft Azure, providing hardware-level isolation for hosted model inference. According to the Cohere product and research blog, the scope of encryption extends from data at rest and in transit to prompts, documents, responses, and model execution during inference; after application payloads are end-to-end encrypted, they can only be decrypted in a confidential environment that has passed remote attestation. Customers can cryptographically verify the hardware, firmware, and software configuration before sending sensitive data, and the client will refuse to send requests when the security policy is not met. The existing Model Vault API and request format remain unchanged. The feature is currently in public beta.

Cohere launches confidential computing public beta on Model Vault

On October 22, 2026, Cohere launched confidential computing on Model Vault, currently in public beta. The feature runs on NVIDIA GPUs on Microsoft Azure, combining confidential CPU and GPU execution, OHTTP end-to-end encryption of request content, and remote attestation, allowing customers to verify the target environment before releasing sensitive data. Cohere also plans to open-source the confidential inference stack and advance support for NVIDIA Blackwell and B200 GPUs.

Jensen Huang Discusses neocloud Financing and High Cost of Capital

On October 22, 2026, Jensen Huang offered his assessment of neocloud financing and the high cost of capital. He said that Nvidia invests in neoclouds because they are good investments, which is different from an endorsement. He stated that neoclouds have ample customers and are sold out; the only problem is a lack of working capital to buy land, power, buildings, and chips in order to fulfill contracts. He believes that rising borrowing costs are a side effect of enormous AI demand, not a flaw in the business model, and that when demand is high, other problems will be resolved. Currently, the developments supported by qualified factual evidence remain limited to statements made at this point in time, with no subsequent progress yet.

MSI releases 1.3L mini PC EdgeMesa N AI+

On October 23, 2026, MSI released the EdgeMesa N AI+ mini PC, with a chassis volume of 1.3 liters. The product is equipped with the NVIDIA RTX Spark N1X superchip and up to 128GB of LPDDR5X unified memory, with running 120B-parameter large models locally as its primary design goal. The chip integrates a 20-core Grace CPU and a Blackwell GPU with 6144 CUDA cores, and the unified memory pool lets model weights and KV cache share the same physical memory. The report, citing real-world tests of the DGX Spark on the same platform, states that 120B-class models generate at roughly 15 to 59 tokens/s, limited by the 273GB/s memory bandwidth. The official price has not yet been announced; the report estimates it may fall in the 30,000 to 40,000 RMB range. There are currently no further developments.

Procore announces deepening of cooperation with NVIDIA

Procore Technologies, a construction management software company, announced at the Groundbreak 2026 conference a deepening of its cooperation with NVIDIA, bringing NVIDIA Omniverse-based digital twin and simulation capabilities into complex construction projects. Reports indicate that the feature is currently in closed beta; users can switch between Procore's Model Manager and Realism Mode to simulate in a high-fidelity 3D environment the impact of real-world conditions such as extreme weather on construction. The cooperation is mainly aimed at large infrastructure projects such as AI data centers. There is currently no subsequent development supported by evidence.

Mifeng and other co-initiators announce the results of this round of RoboColiseum evaluation

Mifeng Technology, together with nearly 20 institutions including Hugging Face, NVIDIA, and Tsinghua University, co-initiated the embodied evaluation platform RoboColiseum, and announced the results of this round of evaluation on October 20. According to a Leikeji report on October 23, Claude Opus 5.5 ranked first overall among the four flagship models, followed closely by GPT-6 Astra; GPT-6 Astra scored 99.35 on instruction following, while Opus 5.5 performed better on spatial understanding and general manipulation, and consumed the fewest tokens. The platform contains 78 simulation evaluation tasks, with a simulation-to-real correlation of 89.5%, and more than 100 teams have participated since it launched in August.

MiniMax thanks teams including SGLang for supporting M3.1 Flash

On October 19, 2026, MiniMax released the M3.1 Flash model. The company says it costs one-third as much as M3 and is 3 times faster, with a unified API price of $0.15/$0.60 per 1M input/output tokens and cached reads at $0.015 per 1M tokens; it had previously been tested anonymously on OpenRouter under the name Space Bunny, ranking first on the model leaderboard for 10 consecutive days. On October 20, the model launched on platforms including Ollama, GMI Cloud, Cline, and OpenClaw, with Cline offering it for free; third-party testing compared it with Claude Opus 5.5, and both completed a Three.js task on the first attempt. On October 21, Parallel Web Systems' search capability efficiency leaderboard ranked it first. On October 22, MiniMax announced that Inferact had become its API launch partner, and the model launched on Runware. On October 23, MiniMax said SGLang and Miles provided support on the first day of release.

View story:MiniMax releases the M3.1 Flash model

SpaceX launches Super Intelligence partner network

On October 22, 2026, SpaceX launched the SpaceX Super Intelligence partner network, aimed at companies building products, solutions, and services based on Grok. The program is divided into two categories: Transformation Partners and Marketplace Partners. According to xAI News, more than 130 companies have participated in the program during the early access phase, and the Forward Deployed Engineer certification program has trained engineers at more than 30 organizations. Bain & Company, Databricks, NVIDIA, Snowflake, and others are among the first Launch Partners. There are currently no subsequent development facts.