Elorian released its first visual reasoning model, Glimpse, emphasizing that it participates in the reasoning process through image generation and editing; simultaneously launched the DecoyCount benchmark to test whether models truly count rather than relying on visual shortcuts; Glimpse scored 50.48% in an internal single-run test, being the only model exceeding 50%, while GPT-6 Astra, Gemini-3.8-flash, Qwen3.8-max, and Claude-Fable-5.1 scored 43.17%, 32.71%, 28.31%, and 25.22% respectively; all evaluation images and texts were not used for training Glimpse.
Cohere's official blog released a reference architecture for a financial research assistant, combining Cohere Embed (Pro/Fast), Rerank (Pro/Fast), Parse, and AWS services (SageMaker AI, OpenSearch Serverless, Bedrock). It supports page-level retrieval from PDF financial reports, table/chart understanding, multilingual queries, and answer generation with source attribution. The solution emphasizes verifiability (each answer is annotated with page number and content type), cost control (Embed Pro for indexing + Embed Fast for querying, int8 1024-dimensional vectors), and flexible toggle design (Rerank, Context Focus, and Page Image are all configurable options).
Moonshot AI released Kimi K3.1 Pro, the first open-weight large language model at the 3T scale, with a context window of 1 million tokens; simultaneously launched the accelerated version Kimi K3.1 Pro Highspeed, offering a 2–3x speed increase; the model achieved 65.7% on Terminal-Bench 4.0 and 61.2% on Kimi Code Bench 3.0 Long-horizon, improved financial modeling accuracy by 7.2 points, and reduced the hallucination rate from 23.4% to 8.3%; to control cyber offense and defense capabilities, relevant data was removed during training, causing the xdev-bench-hard score to drop from 90.5% to 81.0%; full model weights will be open-sourced on November 4, 2026.
The official AMD blog introduces how to optimize the inference efficiency of World Action Models (WAMs) on AMD GPUs (MI355X/W7900) to support closed-loop robot control.
Cloudflare has enhanced Cache Rules on its global edge network to natively support three handling methods for HTTP Vary response headers: normalization, pass-through, and cache bypass, helping operators avoid cache thrashing; this feature is now available to all plans including Free, Pro, Business, and Enterprise.
On October 22, 2026, the US Treasury Department released a proposed rule to abolish the computing power threshold in AI investment reviews, upgrading restrictions on AI investment in China from tiered management to a principled prohibition. The article points out that this adjustment reflects three cracks in the original logic of computing power control: DeepSeek trained frontier models with low computing power, shattering the assumption that 'computing power = capability'; the scale of gray market smuggling is massive (for example, a $160 million H100/H200 network was dismantled); and companies such as ByteDance, Alibaba, and Tencent legally bypass controls by remotely accessing computing power through overseas data centers. Meanwhile, it analyzes NVIDIA's lobbying motivations and the reality of protected markets behind domestic chip price hikes and explosive performance, noting that controls are moving toward passive adjustments due to misaligned interests among multiple parties.
On September 8, OpenAI released a 166-page proof claiming to resolve the existence of singularities in the Navier-Stokes equations (in cases allowing smooth external forces), generated collaboratively by approximately 10,000 AI agents over 88 hours. The proof contains about 130 billion tokens and was translated into Lean language by GPT-6 Astra for machine verification. As of October 26, mathematicians such as Scott Armstrong noted that experts have largely understood its core idea: using the fluid's own instability to spontaneously generate oscillatory pulses to construct external forces that meet smoothness requirements. This method is similar to convex integration but represents a novel implementation. Currently, there are some discrepancies between the original manuscript and the Lean formalized version; the Clay Mathematics Institute still lists it as 'Active', and OpenAI has stated it will not claim the $1 million prize.
As reported by Xinzhiyuan, independent developer Pieter Levels announced he has canceled almost all work software subscriptions, switching to using AI to instantly generate small tools that meet specific needs, such as image scaling services and rent expense calculators; he claims this saves approximately $25,000 per month, and points out that the "era of personal software" is arriving—users state their needs first, and AI then creates the tools; meanwhile, he still pays for infrastructure services such as domains, cloud storage, and AI models.
ModelBest has open-sourced two models: MiniCPM-OCR 1.0 (1.3B parameters) and MiniCPM-OCR Jev. The former achieves SOTA results across 4 tests—Real5-OmniDocBench, Wild-OmniDocBench, OCRBench v2, and CHAOS-Bench—and supports Chinese-English bilingual, PDF, and document parsing under complex acquisition conditions. The latter is an independent decision-making model that does not generate text but only provides millisecond-level confidence scoring, achieving an AUROC of 93.6, and can be plug-and-play adapted to any OCR pipeline.
A team from the University of Tokyo proposed the Multimodal Interactive Deception Assessment benchmark MIDA, constructing 2360 samples containing video, audio, dialogue history, and rules based on the game 'One Night Ultimate Werewolf', with utterances labeled as true/false/neutral; models such as GPT-4o and Qwen2.5-VL-7B achieved an accuracy as low as 27.6% when forced into binary classification, with an F1 score of 0; the study found that dialogue history is far more critical than single-frame visual cues, and designed SoCoT and DSEM modules to enhance social reasoning capabilities; the paper has been accepted by the EMNLP 2026 main conference, and the code and dataset have been open-sourced.
Amazon Web Services (AWS) has open-sourced Strands Decider 2B (a 2 billion parameter decision model based on Qwen3.5-2B with modified head structure, supporting option scoring and confidence output, providing weights, data, and training scripts) and Strands Box (a kernel-level sandbox in developer preview stage, combined with the Dogwood policy language, supporting policy-based control of Shell/Python/MCP/file/network operations).
On October 27, Anthropic released a research report stating that its unpublished internal Claude research model discovered vulnerabilities in Chrome, Safari, Android, Windows, and Linux, and constructed 5 runnable attack chains; among them, Chrome reported 62 critical or high-severity CVEs, and the Linux vulnerability CVE-2026-98432 can achieve zero-click root privilege escalation within 53 minutes. All vulnerabilities have been submitted and fixed according to coordinated disclosure processes.
Microsoft disclosed actual Azure revenue of $3.32 billion for the July–September 2026 quarter for the first time, representing a 48% year-over-year increase and exceeding analyst expectations of $3.23 billion; this figure was slightly higher than Google Cloud's $3.07 billion during the same period, but its growth rate trailed Google Cloud's approximately double year-over-year increase; Azure is now Microsoft's largest and fastest-growing product line, accounting for approximately 36% of its total quarterly revenue of $9.23 billion.
SK Hynix reported Q3 revenue of 104.6 trillion KRW (approx. $7.8 billion), up 328% year-on-year; operating profit of 83.3 trillion KRW (approx. $6.55 billion), up 632% year-on-year, with an operating margin of 80%. The company said its DRAM and HBM capacity is essentially sold out through 2029, expects the memory supply-demand imbalance to persist until 2030, and has signed long-term agreements with 14 key customers covering more than 50% of 2027–2028 capacity.
At the Select event, Supabase announced several new capabilities: support for running a local backend without Docker (alpha), declarative database schema management (inspectable and editable via Git), Supabase Compute (private alpha, placing computation adjacent to the database), OrioleDB and Multigres (high-availability Postgres solutions), and the acquisition of Turso to support horizontal scaling.
AWS has released an enterprise-grade agent platform blueprint built on Amazon Bedrock AgentCore. It supports a federated operating model, including a platform control plane and repeatable Workstream deployment units. The platform offers capabilities such as multi-account topology, managed inference and tool gateways, runtime evaluation gating, and zero-residue retirement, accompanied by a GitHub reference implementation.
AWS released a tutorial demonstrating how to use the official Ray Train Deep Learning Container (DLC), KubeRay, and Elastic Fabric Adapter (EFA) to perform distributed fine-tuning of the Qwen2.5-1.5B model on Amazon EKS; the solution is based on the pre-built image public.ecr.aws/deep-learning-containers/ray:train-ml-cuda-v1.1, eliminating the need to build custom containers. It uses Ray Train to implement FSDP training across 8 NVIDIA L4 GPUs (on 2 g6.12xlarge instances), with all configurations (network, GPU, Python packages, etc.) injected externally via environment variables and manifests.
In Part 3 of QA Studio (an agentic QA automation reference solution based on Amazon Nova Act), AWS released three new capabilities: Applications as test attribution containers, supporting the organization of use cases and test suites by team; a health metrics dashboard, recording execution results via CloudWatch and providing pass rates, failure counts, health scores (0–100) with volatility penalties, and trend charts; and team-level user management, implementing five role levels (Admin/Team Lead/Editor/Executor/Viewer) and cross-team shared access (including permission ceiling rules) through Amazon Cognito.
PyTorch officially released version 2.15.0, with major updates including: the NCCL backend now defaults to a rewritten implementation and adds a flight recorder; torch.distributed supports one-sided transfers and the NIXL backend; FSDP2 and pipeline parallelism support finer-grained memory and communication control; improved 4-bit (NVFP4) LLM decoding performance on NVIDIA Blackwell GPUs; new FP8 support and faster decoding paths for Apple Silicon; added ROCm 10.0 and CUDA 13.4 wheels; minimum Python support raised to 3.11; MPS backend requires macOS 15 or higher; CUDA wheels no longer depend on the cuRAND library; numerous backward-incompatible changes, deprecations, new features, improvements, and bug fixes.
Artificial Analysis released Intelligence Index v5, significantly restructuring the benchmark: 60% of the weight is based on private test sets or hidden answers. In the coding weight, FrontierCode 2.0 accounts for 15%, with all 132 tasks sourced from real open-source repositories; scoring is based on whether maintainers are willing to merge PRs. DeepSeek V4.1-Flash 1017 (max) ranks first among published weighted models on FrontierCode 2.0 with a score of 28%. v5 also introduced Progress Index v1 for comparing model capabilities over time.
Google DeepMind has released Gemini 4 Carbon, claiming it is the strongest model currently available, with cutting-edge cybersecurity and long-horizon task capabilities. It achieved SOTA performance on unprotected benchmarks including SEC-bench Pro (88.2%), ExploitGym (54.8%), and CWE-bench (71.0%). It supports long-horizon agentic workflows and has reduced high-risk misalignment alerts by nearly 80% in internal team testing. Access is being gradually opened to some cybersecurity professionals through the Fairwind program, priced identically to Gemini 4 Argon at $2 per million input tokens and $10 per million output tokens.
Google has launched the Connected Apps platform, enabling US developers to integrate third-party apps with Gemini and Google Search via the AppFunctions API and MCP protocol. It supports choosing to launch on Search only, Gemini only, or both simultaneously. Currently in public preview, over 50 apps are available in Gemini, including Adobe, Experian, and Peloton. Instacart, Canva, and YouTube Music have already integrated with Search, while Calm is set to launch on both platforms soon.