Headlines

1. OpenAI unveils GPT-5.5 “Atlas” with persistent memory and multi-agent orchestration
OpenAI released GPT-5.5 Atlas, a new flagship model that maintains a 1-million-token working memory across sessions and can spawn up to eight parallel sub-agents for long-horizon coding, research, and enterprise automation tasks. The model scores 72.1% on SWE-bench Verified and 86.3% on GPQA Diamond, with API pricing set at $15 per million prompt tokens.
Source: OpenAI Blog

2. EU begins first general-purpose AI compliance audits under the AI Act
European regulators today launched systematic audits of frontier general-purpose AI models, starting with mandatory transparency reports for systems trained above 10^25 FLOPs. The European Commission also announced a fast-track certification lane for open-weight models that disclose training-data provenance and safety evaluations.
Source: TechCrunch

3. Nvidia and TSMC enter volume production on 2nm “Verde Ultra” AI accelerator with HBM4e
Nvidia’s Verde Ultra, built on TSMC’s 2nm N2 process, adds 192GB of HBM4e memory and fifth-generation NVLink, delivering 2.3 exaflops of sparse FP8 compute in a single rack. The first orders were reported from hyperscaler cloud divisions, with general availability slated for Q1 2027.
Source: Reuters

Model Releases

4. Google releases Gemma 3.7 “Scout” open-weight family with native tool routing
Google’s Gemma 3.7 Scout family—available in 2B, 8B, and 27B parameter sizes—features a 1M-token context window and a new “Tool-Select” router that improves agentic benchmark performance by 19% over Gemma 3. The models use 16 trillion tokens of multilingual pre-training data and include “stop-and-verify” hooks for external function calls.
Source: Google DeepMind Blog

5. Mistral AI ships “Large 3.1” with adaptive retrieval and long-hallucination guard
Mistral Large 3.1 adds a Mixture-of-Experts layer that dynamically routes to learned retrieval heads, reducing citation errors by 41% on LegalBench and enabling 512K-token test-time rollouts. The model is available via La Plateforme and as a “hybrid on-prem” deployment package under its enterprise license.
Source: Mistral AI

Industry

6. Microsoft Azure AI Foundry launches Federated Agent Mesh for cross-tenant governance
Azure AI Foundry’s new Federated Agent Mesh lets enterprises deploy AI agents across tenant boundaries with unified audit logging and a policy engine keyed to EU AI Act requirements. Early adopters cite one-click traceability for every tool call, data read, and decision made by a fleet of up to 10,000 agents.
Source: Microsoft News Center

7. Groq and Lenovo debut “Rayliner” 1U inference server with 128 LPUs per node
Groq and Lenovo unveiled the Rayliner, a water-cooled 1U chassis packing 128 tensor-streaming LPUs for 1.4 PFLOPS of native-bf16 inference per node. Rayliner is optimized for Llama 3.1-class and Gemma-3.7-class workloads, claiming 1.8 ms median time-to-first-token for 8B models at batch size 64.
Source: HPCWire

Research Highlights

8. DeepMind’s “Latent Chain-of-Thought 2” shows iterative self-correction without verbose scratchpads
In a new paper, DeepMind researchers introduce LA-STEP, a latent reasoning algorithm that performs backtracking and refinement in hidden state space instead of text-based chain-of-thought. On ARC-AGI-2 and MATH-500, LA-STEP achieves stronger scaling with 37% lower per-token inference cost compared with explicit CoT baselines.
Source: arXiv

9. Stanford–Berkeley study identifies “test-time compute saturation” threshold and proposes entropy-based early stopping
Analyzing 40 reasoning models, researchers found that repeated sampling and long reasoning paths show sharply diminishing returns after roughly 256 generated tokens per problem. The team’s proposed internal-entropy early-stop rule preserves accuracy while cutting average inference compute by more than 3× across FrontierMath and GPQA subsets.
Source: arXiv

Compiled from multiple media reports.