AI Today: UltraQuant AMD and Liquid AI LFM2.5-VL

State of knowledge as of 2026-09-26. Source material gathered between 2026-09-25 and 2026-09-26.

Today's summary covers new developments in language model optimization and cloud services. We present the results of implementing OpenAI tools in fleet management, an update to the GeForce NOW service, and technical improvements in vision-language model architectures. We analyze the UltraQuant technique that accelerates processes on AMD Instinct processors, as well as a new framework for managing relationships in multi-task models used in recommendation systems.

Highlights Today (TL;DR)

• Proaction: The use of Codex, GPT-Live-1, and GPT-6 Astra increased sales by 60% in fleet management.
• GeForce NOW: The platform released the game Control Resonant and announced support for laptops featuring Gemini Intelligence.
• Liquid AI: The LFM2.5-VL-DSpark model utilizes speculative decoding, increasing token generation speed by more than threefold.
• UltraQuant: An optimization method on AMD Instinct processors increases tokens per second by 29% while compressing KV-cache to 4-bits.
• Relation Learning Framework: A new label distribution approximation method in multi-task systems improved recommendation accuracy on YouTube.

Proaction deploys OpenAI tools in fleet management

Proaction reported on September 25 that it is using Codex, GPT-Live-1, and GPT-6 Astra to build, operate, and sell modern fleet management systems. Thanks to the integration of these solutions, the company increased sales by 60% and saved over 75 hours of work.

For AI builders: companies offering specialized business operating systems can check how language models support the automation of sales processes. The announcement does not specify the licensing costs of these solutions or whether they require technical infrastructure adjustments on the client side – these details should be verified before implementation.

GeForce NOW supports Control Resonant

The GeForce NOW platform announced on September 24 the release of Control Resonant in the cloud. This title from Remedy Entertainment allows players to take on the role of Dylan Faden, who uses supernatural abilities to fight threats in an alternative version of Manhattan.

Ultimate subscribers can run the title with GeForce RTX 5080-class performance, utilizing NVIDIA DLSS 4 technologies, ray tracing, NVIDIA Reflex, and cinematic streaming up to 5K with HDR support. The solution eliminates the need for a local installation of the game, which has a size of 100 GB. Until September 27, a subscription bundle is available where purchasing a 12-month membership includes Control Resonant at no additional cost.

The company also announced upcoming support for the service on Googlebooks, a new category of laptops designed for Gemini Intelligence. These devices feature the Titan C chip and integrated system security. Integration with GeForce NOW will enable streaming games with GeForce RTX performance on this hardware later this year.

Additionally, other titles joined the GeForce NOW library, including DragonSword: Awakening, offering open-world gameplay in an anime style. This game supports devices such as Steam Deck, the Firefox browser, and Amazon Fire TV. Other new additions include Dune: Awakening, Stick It to the Stickman, and LEGO Batman: Legacy of the Dark Knight. First-time users of the platform can test cloud access using a 1-day pass, the cost of which can be deducted from the first month's subscription fee.

For AI builders: cloud processing system architects can analyze high-resolution streaming mechanisms using DLSS 4 technology to optimize network load during remote graphic rendering. The announcement does not provide technical specifications regarding transmission latency under various network conditions or the availability of cloud infrastructure in specific geographic regions, which requires verification before implementing custom solutions based on GeForce NOW.

Liquid AI accelerates vision-language models with LFM2.5-VL-DSpark

Liquid AI released an experimental drafter-type model on September 24, designed to optimize inference for the LFM2.5-VL-3B vision-language model. The solution utilizes speculative decoding, which increases token generation speed with minimal memory overhead and no impact on output quality.

The additional component adds 280 million parameters, representing an 8.9% increase compared to the base 3B model. During the inference process, it uses an architecture with 4 attention layers and block sizes of 8 or 9, depending on the hardware used.

Test results indicate significant decoding acceleration compared to the standard model. On Apple silicon devices (e.g., M5 Max), decoding speed increased from 2.30x to 3.13x depending on the task, with an end-to-end latency improvement of 1.56x to 2.62x. For H100 units, decoding acceleration ranged from 2.4x to 2.66x, with an end-to-end improvement of 1.64x to 2.27x.

The authors note that speculative decoding only accelerates the decoding stage, not the image encoding process or the prefill phase. In situations where these stages dominate the total response time, the gain from using DSpark is smaller. The system supports llama.cpp, MLX-VLM, and SGLang. The models are available on Hugging Face in Safetensors and GGUF formats.

For AI builders: developers deploying vision systems on edge devices can integrate the model with libraries such as SGLang using the runtime parameters specified by the creators. The documentation does not indicate whether the solution supports all natural languages equally or what the licensing restrictions are for commercial use – these must be verified before deployment in a production environment.

UltraQuant on AMD Instinct processors

AMD published details on September 25 regarding the UltraQuant technique, which increases decoding throughput and lowers latency for agentic language models on AMD Instinct GPUs. This solution is dedicated to models operating in long-context environments, such as Qwen3.8-2.4T with MXFP4 weights.

The UltraQuant technique compresses the key-value cache (KV-cache) to 4-bits by leveraging native FP4 support via CDNA4 units in MI355X chips. Unlike traditional compression, this method does not require unpacking data before matrix multiplication operations, as scaling is integrated directly into MFMA instructions. In tests conducted using vLLM (version 0.19.2rc0), AITER (0.1.16.post3), and FlyDSL (0.2.0) libraries under a C=32 load, UltraQuant achieved 29% higher tokens per second and 24% lower latency compared to the fastest 8-bit backend.

Under low-load configurations, performance is comparable to standard 8-bit methods, but UltraQuant's advantage grows as the number of concurrent requests increases. The authors also point to the optional use of a Multi-Token Prediction (MTP) head, which under moderate load (up to C=16) further reduces inter-token latency by 12–22%, while under high load, its deactivation is recommended. Tests on GPQA-Diamond and SWE-bench Lite datasets confirmed that applying UltraQuant does not affect measurable reasoning accuracy of the model.

For AI builders: Engineering teams running agentic systems with long conversational sessions can test this method on AMD Instinct infrastructure. The material does not specify software license requirements or information regarding UltraQuant availability in hardware environments outside the CDNA4 architecture – these aspects require verification before production deployment.

New relation learning framework in multi-task models

According to a publication on September 25, a research team proposed a method for learning relationships between multiple tasks in multi-task models. This solution relies on approximating the joint label distribution through targeted pairwise relationships, avoiding the complexity of modeling the full joint space. The authors tested the method's effectiveness in YouTube recommendation systems, utilizing it in notification modules, the homepage, and the "watch next" feature. Results indicate increased accuracy and user satisfaction. The framework enables transfer learning and improves information extraction. The documentation also includes a workflow template designed to facilitate future deployments in other multi-task systems.

For AI builders: Teams maintaining multi-task recommendation systems can check whether a pairwise relationship-based approach impacts their models' performance metrics. The publication does not specify the hardware requirements necessary to run this framework or provide license information for using the proposed workflow template.

Frequently Asked Questions (FAQ)

By how much did decoding performance increase in the Liquid AI model on Apple M5 Max chips?

For Apple silicon devices such as the M5 Max, decoding speed after applying the LFM2.5-VL-DSpark model increased from 2.30x to 3.13x depending on the task performed. End-to-end latency improvement ranged from 1.56x to 2.62x.

What benefits does the UltraQuant technique bring on AMD Instinct processors?

The UltraQuant technique compresses the KV-cache to 4-bits, achieving 29% higher tokens per second and 24% lower latency compared to the fastest 8-bit backend. This method does not require unpacking data before matrix multiplication operations.

Sources

• Proaction boosts sales 60% and saves 75+ hours with Codex — 2026-09-25
• Contain the Chaos: ‘CONTROL Resonant’ Launches on GeForce NOW — 2026-09-24
• Accelerating vision-language models with LFM2.5-VL-DSpark — 2026-09-24
• UltraQuant on AMD Instinct: More Efficient Agentic Serving for Qwen3.8-MXFP4 — 2026-09-25
• Learned Cross-Task Relationships in Multi-Task Models — 2026-09-25

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