AI Today: NVIDIA RTX Spark and Gemini 4 Argon

State of knowledge as of 2026-10-14. Source material gathered 2026-10-14.

Today's review focuses on the new NVIDIA RTX Spark mobile hardware and the release of the Gemini 4 Argon language model. The summary also covers video tool updates from Higgsfield, technical parameters of the LFM2.5-24B-A2B model from Liquid AI, and NVIDIA's latest investments in the European tech startup ecosystem. Each of these solutions brings significant parameters to local AI processing, code optimization, and advanced multimedia generation.

TL;DR

• NVIDIA RTX Spark: A mobile chip with 128 GB of memory and 1 petaflop of AI computing power debuting on October 16.
• Gemini 4 Argon: Google's new model focused on engineering and science, available to developers via AI Studio.
• Higgsfield: A platform update allowing the generation of video materials lasting over 10 minutes.
• LFM2.5-24B-A2B: An MoE type model from Liquid AI, optimized for fast performance on local devices.
• NVIDIA Investments: Financial support for French startups dealing with agentic computing across various industrial sectors.

NVIDIA RTX Spark

Nvidia has announced the launch of the RTX Spark chip, which combines the capabilities of a gaming PC, a creative studio, and a personal AI supercomputer. Laptops equipped with this chip, offering up to 128 GB of unified memory and 1 petaflop of AI computations, will go on sale on October 16.

The device is designed to support local AI agents, for which updated tools and models have been made available. Among the novelties is the Nemotron 3.5 Bolt 74B model, designed specifically for AI agents, which will launch on October 15. The hardware supports numerous programming frameworks such as PyTorch, Python, and CUDA libraries. Cooperation with partners like Adobe is expected to bring a twofold increase in performance in applications such as Photoshop and Premiere.

In terms of security, RTX Spark utilizes NVIDIA OpenShell and Microsoft Execution Containers (MXC) technologies to ensure agent environment isolation. Users will also gain access to the NVIDIA Switchyard library, used for intelligent routing of queries between local and cloud models. Furthermore, the NVIDIA PAIR tool is exiting beta, enabling the distribution of AI workloads across multiple devices on a local network with support for llama.cpp.

For AI builders: A person deploying local AI agents can test a configuration based on the new Nemotron 3.5 Bolt 74B model and Switchyard query routing tools. Before making a purchase decision, one should determine whether this model has sufficient Polish language support and what the actual electricity costs are when fully utilizing 1 petaflop of compute on a mobile device. The description also does not clarify whether the MXC security mechanisms are fully compatible with corporate operating systems managed by IT policies.

Google Introduces Gemini 4 Argon

Google released the new Gemini 4 Argon language model on October 13, designed for advanced engineering, scientific, and business tasks. The model is currently available globally for developers and enterprises via the API in AI Studio and the Gemini platform, as well as for Google AI Ultra and Pro subscribers in selected services like the Gemini app and AI Mode in search.

Gemini 4 Argon stands out for its precision in multi-step tasks and ranks first in the Vals Index, which evaluates performance in the financial, legal, and programming sectors. According to the manufacturer, the model achieves high efficiency in working with complex code, cybersecurity, and business process optimization. Internal tests used it, among other things, to refactor 6 million lines of code in Chromium and accelerate calculations in quantum chemistry packages by a factor of 2.3x. In infrastructure, using the model allowed for a 2.85x speedup in inference on A100 units.

The pricing of the model is set at 2 dollars per million input tokens and 10 dollars per million output tokens, with an option for a 95 percent discount on input tokens in cache mode.

For AI builders: Software engineers and data analysis teams can test the model within the AI Studio environment for the automation of multi-step tasks. The announcement does not state whether the model has Polish language support comparable to English, what the requests per minute (RPM) limits are, or if detailed license terms for commercial deployments outside the Google environment are available — this needs to be checked before starting work.

Higgsfield Introduces Video Tool Updates

Higgsfield released new tools for creators on October 13 that enable the generation of longer video materials and precise graphics. The platform now allows for the creation of videos lasting over 10 minutes, representing an increase from the previous 1-minute limit. The new features rely on GPT-6 Astra, GPT-Image-2.5, and GPT-Live technologies, with the latter intended to support the film creation process using voice.

For AI builders: video creators and animators can test the performance of generating long sequences in the new model. The announcement lacks information on Polish language support availability, precise costs per minute of generated footage, and details regarding commercial usage licenses for videos created using these technologies.

LFM2.5-24B-A2B Model from Liquid AI

Liquid AI presented the LFM2.5-24B-A2B model on October 13, designed to handle fast and private AI agents directly on user devices.

The model's architecture is based on 24 billion total parameters, of which only 2 billion parameters are active during the processing of each token. Such an MoE (Mixture of Experts) architecture helps reduce the computational costs of a single reasoning token. The model uses 40 layers, 64 experts per block, and the GQA mechanism. The context window has been increased to 131 thousand tokens, and the vocabulary expanded to 128 thousand tokens. The training process covered 33 trillion tokens, and post-training incorporated MOPD (Multi-domain on-policy distillation) and Agentic RL methods.

In tests, the model achieved a score of 48.09 on the LiveClaw benchmark. Locally on an Apple M5 Max processor, it achieves a speed of 184 tokens per second, while on a single NVIDIA B300 chip, this speed reaches over 21 thousand tokens per second. The tool shows an advantage in tool-use and instruction-following tasks, while in tasks requiring a large knowledge base or advanced programming, it yields to larger models. Support for the model is provided by llama.cpp, MLX, vLLM, SGLang, and ONNX platforms.

For AI builders: A developer building local agentic systems can implement this model using the transformers library or combine it with popular environments like the Hermes Agent. The documentation does not specify cloud operating costs at scale nor does it address the exact level of Polish language support, which must be verified before integration in a production environment.

NVIDIA Invests in French Agentic Startups

NVIDIA announced new investments on October 13 in the ecosystem of French startups dealing with agentic computing. Support was received by Doctolib, Entalpic, Genesis AI, and Pigment, which are developing technologies for healthcare, materials engineering, robotics, and business planning. These actions fit into the corporation's broader commitment in Europe, where nearly 10 billion dollars have been channeled into technological development.

The French ecosystem currently utilizes NVIDIA infrastructure to create agentic solutions in sectors such as finance, public administration, the space industry, and scientific research. For example, Bull plans to expand AI server production based on NVIDIA system Vera Rubin Ultra NVL72, and the Altair-1 satellite uses NVIDIA Jetson AGX Orin modules along with Mistral models for autonomous Earth observation. Meanwhile, L’Oréal is integrating the NVIDIA ALCHEMI solution to accelerate research processes, and public organizations such as LINAGORA and Docaposte are deploying solutions based on NVIDIA Nemotron models under the CANUT agreement. In the field of quantum technologies, Alice & Bob and Quandela are using the NVQLink platform to optimize cooperation between QPUs and GPUs.

For AI builders: Engineering teams considering the deployment of Nemotron models or Cosmos infrastructure can check the availability of these solutions for their own sovereign data environments. The announcement does not state the license for using model weights in commercial applications or detailed maintenance costs for the NVL72 infrastructure — these issues require confirmation through direct contact with the vendor.

Frequently Asked Questions (FAQ)

What are the capabilities of the Gemini 4 Argon model?

The model exhibits high precision in engineering, scientific, and business tasks. In internal tests, it was used to refactor 6 million lines of code in Chromium and accelerated calculations in quantum chemistry packages by a factor of 2.3x. It ranks first in the Vals Index evaluating performance in the financial, legal, and programming sectors.

What distinguishes the architecture of the LFM2.5-24B-A2B model?

The model relies on an MoE architecture with 24 billion parameters, of which only 2 billion are active during the processing of each token. The design uses 40 layers, 64 experts per block, and the GQA mechanism, which allows for reduced computational costs while expanding the context window to 131 thousand tokens.

Sources

• All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark — 2026-10-13
• Gemini 4 Argon: Built to solve your hardest problems — 2026-10-13
• Gemini 4 Argon is now available for developers. — 2026-10-13
• Higgsfield takes AI video from 1 minute to over 10 — 2026-10-13
• Run fast, private agents on-device with LFM2.5-24B-A2B — 2026-10-13
• NVIDIA Announces New Investments in France’s Startups — 2026-10-13

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