AI Today: NVIDIA DGX Spark and OpenAI Astra Ultrafast

State of knowledge as of 2026-10-03. Source material collected between 2026-10-02 and 2026-10-03.

Today's review covers key updates in the hardware and model sectors. NVIDIA introduces the DGX Spark unit for local AI computing, while OpenAI increases the performance of the GPT-6 Astra Ultrafast model. In the open-source domain, the AstaBrief 8B tool and the Olmo-core 3 training infrastructure debut. The summary concludes with information about a new training program for deployment engineers launched by Anthropic.

TL;DR

• NVIDIA DGX Spark: A local computing unit based on the Grace Blackwell architecture, supporting models up to 100 billion parameters.
• GPT-6 Astra Ultrafast: A new OpenAI model mode utilizing NVIDIA Blackwell processors, delivering eight times faster token generation.
• AstaBrief 8B: An open-source model optimized for rapid generation of scientific reports based on literature.
• Olmo-core 3: Infrastructure from Allen Institute for AI enabling the training of Mixture-of-Experts models at a scale of up to one trillion parameters.
• Claude Frontier Academy: An Anthropic educational program aimed at training 10,000 engineers in deploying Claude models within enterprises.

NVIDIA DGX Spark 64GB

NVIDIA announced the market release on October 23 of the DGX Spark device equipped with 64 GB of unified memory, designed for building and scaling local artificial intelligence systems. The hardware is based on the Grace Blackwell architecture and includes the full NVIDIA software stack, including the DGX OS.

The platform enables running models sized up to 100 billion parameters directly on the device, allowing work on AI agents, inference, model fine-tuning, and data analysis without cloud usage. The kit includes a ConnectX-7 network card, which allows connecting two units into a cluster using the NVIDIA Sync Cluster Assistant. In this configuration, users gain 128 GB of unified memory and support for models up to 200 billion parameters, which in performance tests with the Qwen 3.8 27B model translated into a 1.7x performance increase compared to a single system.

The device supports popular tools and frameworks, such as NVIDIA Agent Toolkit, Ollama, vLLM, PyTorch with CUDA, and CUDA-X AI libraries. The hardware will be available from manufacturers Acer, ASUS, Dell, Gigabyte, HP, and MSI, starting at a price of 4,999 USD. By the end of the month, the release of the NVIDIA Sync Model Launcher is planned, a tool facilitating model configuration in single- and multi-node environments.

For AI builders: Systems engineers and researchers working with local agents can use this hardware to test models in an environment isolated from the external network. The announcement does not provide information on the device's availability in different regions of the world, detailed licensing conditions for the software bundled with the hardware, or the maintenance costs of infrastructure in multi-node clusters – these data should be verified with selected manufacturers before purchase.

OpenAI releases GPT-6 Astra Ultrafast

OpenAI released the GPT-6 Astra Ultrafast model on October 1, utilizing NVIDIA Blackwell graphics processors to increase data generation performance. The new mode is currently available via OpenAI API and for ChatGPT Work and Codex users.

GPT-6 Astra Ultrafast offers token generation up to 8 times faster than the Astra Standard version. Utilizing the NVIDIA Blackwell architecture allows for the creation of high-performance kernels, translating into lower latencies in operational loops such as writing code, using tools, or testing solutions by agents. OpenAI also uses its own models to optimize inference software on NVIDIA infrastructure, allowing for continuous improvement of model response speed and efficient management of computing resources. The programmability of the NVIDIA platform allows teams to reuse infrastructure in training, inference, and reinforcement learning processes.

For AI builders: Creators of automatic programming agents can check response generation times within edit, test, and debug cycles. The announcement does not provide a detailed pricing list for using the Ultrafast mode or information on token usage limits, which should be verified in the technical documentation before deciding to migrate a production environment.

Release of the AstaBrief model

Organizers of the Asta platform published the AstaBrief 8B model as open-source on October 2, designed for rapid generation of scientific reports based on literature. The tool was designed to shorten documentation creation time by generating a full report in a single pass, serving as an alternative to the previous sectional workflow based on external models.

According to the creators' measurements, implementing AstaBrief 8B shortens the average report generation time to 51.1 seconds, which is about 3.5 times faster compared to the Thinking mode using proprietary models (178.5 seconds). The model training process was based on 47,000 training examples filtered from 90,000 research queries for quality and relevance to scientific literature. The model was trained using supervised fine-tuning (SFT) and Direct Preference Optimization (DPO), forgoing more costly methods based on reinforcement learning. Users can use AstaBrief in Fast mode directly on the Asta platform or download model weights to run it on their own infrastructure, enabling the processing of confidential or unpublished data.

For AI builders: Teams working on scientific literature synthesis systems can analyze the shared workflow for generating reports from their own PDF files. The announcement does not specify whether the model supports languages other than English, what the actual infrastructure maintenance costs are with a high number of queries, or exactly what licenses cover the shared weights and training data – these issues require independent verification before deployment.

Olmo-core 3

Allen Institute for AI released the Olmo-core 3 training infrastructure on October 1, designed for training large Mixture-of-Experts (MoE) models. The system aims to enable scaling model learning processes to the level of one trillion parameters while maintaining computational efficiency.

Compared to previous implementations based on Fully Sharded Data Parallelism (FSDP), the new version uses an approach based on Distributed Data Parallelism (DDP), allowing experts to be kept in GPU memory and eliminating repetitive weight gathering. In tests conducted on eight NVIDIA B300 graphics processors, a model with 47 billion parameters achieved a throughput of 52,000 tokens per second per GPU, 2.7 times higher than the previous solution.

Olmo-core 3 combines expert parallelism, pipeline parallelism, and distributed optimizer techniques, limiting the need to store the full model and training state in the memory of a single device. The solution supports the MXFP8 number format, which in tests increased throughput by about 21% compared to the BF16 format while lowering peak memory demand from 103 GiB to 95 GiB. Research work tested configurations scaling up to 2.38 trillion parameters. The technical documentation also includes analyses concerning the "token gerrymandering" phenomenon and challenges related to load balancing experts during training.

For AI builders: Researchers dealing with MoE model scaling can check the source code and interactive demo on the Hugging Face platform. The published report does not specify whether the tool has native support for hardware from manufacturers other than NVIDIA, or what the energy costs are for long-term training of trillion-parameter models – this information should be verified before starting work on your own infrastructure.

Anthropic launches Claude Frontier Academy

Anthropic invested 100 million dollars to launch the Claude Frontier Academy, aiming to train 10,000 Frontier Deployed Engineers (FDE) by the end of 2027. According to the October 2 announcement, the program aims to address the shortage of skilled workers capable of deploying Claude-based solutions inside large enterprises.

Participants of the first cohort were engineers from companies such as Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. The educational model is modeled on medical practices: participants learn from practitioners, work on realistic scenarios, and undergo an evaluation process before receiving certification.

The training process consists of several stages. The program begins with a few days of stationary classes with Anthropic engineers, during which participants go through a simulated deployment – from use case selection, through security review, to system handover to operations. After passing the practical exam, participants receive the Claude Resident Engineer badge and begin a 12-week residency. During this time, they lead a real project utilizing Claude in their organization, supported by Anthropic experts. Completion of the residency and re-evaluation result in being awarded the Claude Frontier Deployed Engineer badge.

Participation in the program is possible exclusively by nomination. Sessions currently take place in San Francisco, New York, and London.

For AI builders: The program is aimed at practicing software engineers who have experience working with LLMs and can deploy them in business environments. Decision-makers in technical teams should contact their Anthropic customer account manager to verify qualification criteria for nomination. The announcement does not contain information about the costs of participation in the training for individual organizations or the availability of training materials for companies outside the partner network.

FAQ

What capabilities does connecting two NVIDIA DGX Spark devices offer?

Thanks to the ConnectX-7 network card and NVIDIA Sync Cluster Assistant software, users can connect two units into a cluster. This provides 128 GB of unified memory and support for models sized up to 200 billion parameters, which in tests translates into a 1.7x performance increase compared to a single system.

Who can participate in the Claude Frontier Academy program?

The program is aimed at engineers with experience working with LLMs, working in large organizations. Participation is possible exclusively by nomination, and training sessions currently take place in San Francisco, New York, and London. Interested parties should contact their Anthropic customer account manager.

Sources

• NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI — 2026-10-02
• How NVIDIA GPUs Help Accelerate OpenAI’s GPT-6 Astra Ultrafast — 2026-10-01
• Open-sourcing AstaBrief, the fast report-generation model in Asta — 2026-10-02
• Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs — 2026-10-01
• Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap — 2026-10-02

--- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Own summaries, not copies. Pre-human-edit version.