Google Cloud published a reference architecture for building AI agents that coordinate long-running, GPU-heavy scientific tasks like protein folding and compound screening, which break standard single-LLM-agent setups due to timeouts and context overflow. The design separates a cognitive orchestrator (Gemini on Agent Runtime) from a GPU execution layer (NVIDIA NIMs like ESMFold and GenMol on GKE) and wraps both in an enterprise governance plane with agent identity, a gateway, a registry and Model Armor. It also lays out best practices such as passing bulky data by reference instead of inline, streaming progress updates, and gating irreversible actions behind human approval.
Reflection, a new US AI lab, has launched Beam, its first model, as an open-weight mixture-of-experts system with 501B total parameters (23B active per token) and a 1M-token context window, released under Apache 2.0. The model is built for coding and agentic use cases, ships with a custom Rust-based inference harness, and has day-0 support in transformers, vLLM and SGLang. Reflection reports strong results on coding, reasoning and tool-calling benchmarks across BF16, FP8 and NVFP4 checkpoints.
NVIDIA, working with IIT Madras's AI4Bharat lab, is releasing Nemotron-pretraining-Indic-v1, a roughly 2.8 trillion token synthetic pretraining corpus covering 22 Indian languages, along with the open GPU-accelerated pipeline used to generate it. The corpus includes native-script data for all 22 languages and romanized text for twelve, built by translating curated English source material (web text, math, code, Q&A) using NeMo Data Designer and quality-checked with an MQM-inspired rubric. NVIDIA already used a 10-billion-token Indic subset for continued pretraining of Nemotron 3.5 Super VL, which showed measurable gains on Indic benchmarks.
IBM has released Granite Switch 4.1 in 8B and 30B sizes, a modular system that runs chat, decision-making and speech transcription from one hosted endpoint instead of separate models. The update adds a general-purpose Decision adapter, a specialized Guardian adapter for safety-related classifications, and an embedded ASR model, Granite-Speech-5.0-Turbo-CTC, which tops the OpenASR Leaderboard. Despite bundling three capabilities, the combined model uses roughly the same VRAM as the base model alone.
Meta published Muse Spark Open, an open-weight mixture-of-experts model with about 816 billion total parameters and 42 billion active per token. It accepts text, images, audio and video as input and generates text, with a context window of up to 1 million tokens. Meta positions it as a coding-optimized, agentic model built on top of the smaller Muse Glimmer architecture, calling it the most capable US open-weights model released so far.
Apple's MLX team published benchmarks showing large open-weight models — including a near-trillion-parameter MoE model called Muse Spark Open — running on Mac Studio hardware, both on a single machine and distributed across up to four Mac Studios using a new RDMA-over-Thunderbolt driver. The new M5 Ultra chip, with dedicated GPU Neural Accelerators and 1.2 TB/s memory bandwidth, delivers up to 4.6x faster time-to-first-token and 1.6x faster generation speed versus the previous M3 Ultra. For models too large to fit on one machine, MLX now supports pipeline and tensor parallelization across multiple Macs connected via Thunderbolt RDMA, letting a four-Mac-Studio cluster generate over 36 tokens per second on an 816B-parameter model.
Swiss AI released Apertus 1.5, a continued pretraining of its open Apertus 1.0 models (8B and 70B, Apache 2.0) with native image and experimental audio understanding, a 4x larger context window of 262,144 tokens, an optional thinking mode, and improved instruction-following and tool use. The models were trained on a total of 19T (8B) and 17T (70B) tokens, with full training data, recipes and intermediate checkpoints published for reproducibility. The team also published detailed, standardized evaluation cards and announced early API access from providers like Swisscom and Infomaniak, plus a roadmap for a ~600B-parameter MoE Apertus 2.0.
Researchers describe a system called FlowSpec that converts a creditor's legal call procedure into machine-readable instructions, tool restrictions, training signals and test cases for a voice agent. A 9-billion-parameter model was trained on calls covering nearly 300 randomly generated procedures, then tested on four procedures it had never encountered, including one entirely held-out family of scripts, to see whether compliance behavior transfers.
GitHub replaced its LLM-based pipeline for detecting generic secrets (like passwords and API keys) with a fine-tuned ModernBERT classifier combined with regex candidate generation. The new two-stage system improved detection accuracy while cutting scan time and serving costs dramatically, and it's now fast enough to run during push protection, not just after code is committed.
Liquid AI has published LFM2.5-24B-A2B, a mixture-of-experts model with 24B total parameters but only 2B active per token, designed to run reasoning agents on local hardware. The release expands the context window to 131K tokens, doubles the vocabulary size, and was trained on roughly double the pretraining data of its predecessor. On agentic and tool-use benchmarks the model performs on par with or better than models that have about 50% more parameters, while delivering fast inference speeds on both consumer and data-center hardware.
Orca Security built Discovery AI, an Amazon Bedrock-based agent that turns plain-English security questions into structured queries against its Sonar system, which tracks over 900 cloud resource types. Rather than relying on one large model with the full schema stuffed into every prompt, the production version splits the work across nine specialized stages and multiple foundation models. The result: dramatically lower cost, faster responses, and better accuracy on complex queries, now running live since March 2026.
AWS detailed three kernel-level optimizations for Mixture-of-Experts (MoE) model serving on Hopper GPUs, implemented in collaboration with the vLLM community and shipped in Amazon SageMaker AI's LMI container. The changes target the inefficiencies of fused MoE computation during low-batch decode, which previously left Tensor Cores underutilized. Combined, the optimizations deliver up to 10% higher throughput and up to 9% faster time-to-first-token (TTFT).
NVIDIA has released the full Run:ai platform as open source, adding it as a foundation of the NVIDIA DSX AI Factory Platform for AI workload and GPU orchestration on Kubernetes. The release includes four modular components — KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map — that can be adopted together or independently. Eight ecosystem partners, including Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud, and vCluster, are integrating the platform into their own offerings.
IBM Research used a coordinated multi-agent 'agent factory' to explore hardware design optimizations for a quantum error-correction decoder, comparing proprietary and open frontier models under the same 100-hypothesis search budget. The open-model configuration, run on-premises via IBM's llm-d distributed inference framework, achieved a 52x speedup over baseline for about $26 in inference cost, versus $299 for a proprietary-model configuration that reached a higher 204x speedup. The study highlights that cache-aware routing and efficient serving (98% prefix-cache hit rate, ~$0.11 per million tokens) made the open-model approach dramatically cheaper while still finding valid, compact hardware designs.
Amazon Web Services researchers published a method called 'priming' that converts pretrained Transformer models into hybrid architectures mixing attention and state space model (SSM) layers, reusing existing weights instead of retraining from scratch. The approach cuts pretraining costs by over 99% and lets a 32-billion-parameter hybrid model match its source Transformer's quality while decoding up to twice as fast. Amazon tested three SSM types and released an open-source model zoo, training code, and serving tools under Apache 2.0.
NVIDIA, working with Princeton's Dao AI Lab, built ReplaySSM, a new way to store state in hybrid models that mix state-space (SSM) layers with attention layers. Instead of rewriting the entire state on every token, the method keeps a periodic checkpoint plus a short list of recent inputs, splitting the job of producing output from the job of updating memory. The result is roughly half the memory traffic, decode kernels running up to 1.80x faster at scale, and speculative decoding throughput more than doubling on 120B-class models, all while keeping accuracy close to full precision. The method already ships inside TensorRT-LLM, vLLM, and FlashInfer, and is running in production for GLM-5.3-Flash and Kimi-K3.
Mistral AI has released Voxtral Mini 4B Realtime Arabic, a streaming speech-to-text model fine-tuned for Arabic dialects, Modern Standard Arabic, and code-switching between languages. The 4.4-billion-parameter model processes 16 kHz audio in real time with a configurable transcription delay, achieving an 8.82% average Character Error Rate across seven Arabic benchmarks at a 480ms delay — close to the offline Voxtral Transcribe Arabic model's 7.91% CER. It was co-developed with Morocco's digital transition ministry as part of a sovereign AI partnership and is released under Apache 2.0.
Google released ML Drift, an open-source, cross-platform GPU compute engine for on-device ML inference, replacing the legacy TFLite GPU delegate. It unifies shader handling across OpenGL ES, OpenCL, Metal and WebGPU, adds 5D tensor support, and introduces stage-aware optimizations for LLM prefill/decode. It is already running in production in Chrome, YouTube Shorts, Google Photos, Meet and partner apps from Adobe and Snap.
Cisco Talos' CAIRN research classifies a new malware archetype, 'A3: AI-Analysis Evasion,' where malicious code embeds natural-language text designed to manipulate language models used in automated malware triage. Talos traced the technique across four malware families (FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE) and 84 samples from January 2025 to July 2026, showing it spreading from a simple copy-pasted comment to templated, multi-model-targeted 'template spraying' and fake intimidation messages. Controlled testing with a panel of five local LLMs found the simplest direct-instruction technique shifted model verdicts toward 'benign' in roughly 35% of runs, while more elaborate tricks had little effect or backfired by increasing suspicion.
Google has released the Developer Knowledge API, a structured, machine-readable source for Google Cloud, Firebase and Android documentation aimed at AI agents and developer tooling. Instead of scraping web pages or relying on an LLM's training cutoff, agents and CLIs can now query fresh, Markdown-formatted docs through several interfaces: a gcloud CLI surface, an installable agent skill, client libraries, and a browser-based API explorer. The underlying index is refreshed frequently so retrieved answers stay close to the latest published documentation.
AWS describes a new two-stage access control architecture for Retrieval Augmented Generation (RAG) in Amazon Quick and Amazon Bedrock Knowledge Bases. Instead of relying only on periodically synced access control lists (ACLs), the system now verifies permissions in real time against the authoritative data source (e.g. Google Drive, SharePoint, Confluence) at query time, before passing content to the LLM. AWS says this closes security gaps caused by stale or incorrectly mapped permissions between sync cycles. Mondelēz International, with over 35,000 employees, is cited as a customer using this approach.
Apple researchers, with National University of Singapore collaborators, published a paper on RISED, a new training framework for LLM agents that must learn across many different interactive environments at once. Instead of relying only on scalar reward signals, RISED uses an LLM judge to tag each training rollout with rubrics describing its behaviour, then uses these rubric tags to select training data and to supply extra supervision. The authors report RISED achieves the highest average success rate across environments compared to existing approaches, ranking first or second in every tested environment.
AssemblyAI launched a hosted deployment of the open-source Qwen3.5 4B model on its LLM Gateway, optimized specifically for rewriting raw speech-to-text output into clean, formatted text. On voice rewrite tasks like dictation cleanup and transcript formatting, the model averaged 612ms response time—1.9x faster than GPT-4.1—at 94% lower cost per hour of audio processed. The model is priced at $0.10/$0.50 per million tokens (prompt/completion) and supports only max_tokens, temperature, and stream parameters, with no tool calling.
Google has launched EmbeddingGemma 2, a Gemma 4-based embedding model released under Apache 2.0 that places text, code, images, video and audio in one shared 768-dimensional vector space. It's modular, so developers load only the encoders they need: 270M parameters for text and code alone, scaling up to 740M for every modality. Google reports a 14% jump over EmbeddingGemma 1 on the MTEB Code benchmark while keeping the earlier model's multilingual text accuracy.