AI Daily (November 7, 2026)

Today's Theme: OpenAI raises billions at $1.25T valuation, Murati team launches 594B model

Overview: On November 7, 2026, the global AI industry and technology ecosystem saw several key changes. On the capital and model front, OpenAI completed the first tranche of approximately $11 billion at a $1.25 trillion valuation; former OpenAI CTO Mira Murati's team released Inkling 2 Small with a total of 594 billion parameters, and plans to open weights next week. In infrastructure and regulation, the Ministry of Industry and Information Technology (MIIT) solicited public comments on a unified coding scheme for all categories of robots; Microsoft is reportedly gradually reducing compute rented from Oracle for older-generation chips. Meanwhile, Cloudflare launched a Rust version of the TypeScript compiler, vLLM explored asynchronous RL with zero error alignment, and multiple agent and software-hardware collaboration tools were also unveiled.

27 Selected Items This Issue

Model Releases

1. Thinking Machines Lab releases 594B-parameter MoE model Inkling 2 Small, plans to open weights next week

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, officially released the general-purpose multimodal model Inkling 2 Small, using a MoE architecture with 594 billion total parameters and 17 billion active parameters, and plans to open weights next week. The model scored 31 on third-party evaluation Artificial Analysis intelligence index, and focuses on supporting long-duration autonomous task execution, improving agentic programming, vision, knowledge work, and reasoning. There are also reports that the company is in talks for a new $5 billion funding round led by Accel, with a pre-money valuation possibly reaching $40 billion.

2. Dharma-AI releases 2B-parameter structured field extraction model Dharma Field Extractor

Dharma-AI launched Dharma Field Extractor, a specialized extraction model with only 2 billion parameters, which can extract specified fields from various unstructured documents and output standard JSON. The company claims it has higher accuracy than tested general frontier models at a fraction of the cost.

3. Ricursive Intelligence launches open-source code model Ripple-65B Preview

Ricursive Intelligence released Ripple-65B Preview, an open-weight language model with 65 billion parameters, under the fully open-source Apache 2.0 license. The company says the total R&D cost was $32 million, and it achieved a leading score on the TerminalBench 2.1 programming benchmark among US open-source models, but lagged behind Inkling 2 Small, released around the same time, on other general evaluations.

Compute Hardware

1. Microsoft reportedly phasing out AI cloud compute rented from Oracle

People familiar with the matter say Microsoft is gradually decommissioning and terminating contracts for AI compute rented from Oracle. The Oracle hardware uses older-generation Nvidia chips that no longer meet Microsoft's current internal specifications and compute needs. Microsoft has been phasing out this external compute for several months without renewing agreements.

2. Amazon SageMaker Inference supports Nvidia MIG multi-instance GPU partitioning

AWS announced that Amazon SageMaker Inference now supports Nvidia Multi-Instance GPU (MIG) hardware partitioning. This allows developers to partition a single GPU into multiple isolated instances at the physical isolation level, using two new API fields to deploy different models per partition or allocate new accelerator compute on demand, improving overall GPU utilization.

Research Papers

1. OpenAI releases BioBasicsBench to evaluate agents' wet-lab analysis ability

OpenAI introduced BioBasicsBench, a benchmark set for biological research scenarios, containing 100 standardized scientific tasks. The benchmark is designed to measure AI agents' actual reasoning and analysis performance in understanding raw experimental data, deriving biological rules, and handling routine wet-lab research workflows.

2. arXiv: SwarmEvolve framework uses latent-space swarm communication to accelerate program and algorithm discovery

A study on arXiv proposes the SwarmEvolve framework, which addresses the slow decoding of long chain-of-thought in LLM evolutionary search by having hundreds of lightweight agents share information via KV caches in latent space. Experiments show that on 21 mathematical discovery and scientific computing tasks, the median time to reach SOTA performance is reduced to 3 to 8 minutes, significantly outperforming existing frontier frameworks that take about 60 minutes to reach similar scores.

3. arXiv: Scientist World builds 1000 executable research environments for AI scientists

Researchers released Scientist World, which converts each paper into an executable environment: automatically extracting the paper's open research questions and reproducing its methods as baselines, allowing an AI-Scientist entering the environment to compare against the human authors' results. The team released an initial batch of 1000 validated environments and found that frontier AI-Scientists outperform human researchers in 74% of cases; these environments also contain 6400 hours of research trajectories, which can improve small models' research ideation ability by 158%.

4. arXiv: XStack benchmark evaluates agents' cross-ecosystem code migration ability

Researchers released XStack, an automated evaluation benchmark for machine learning codebase migration, focusing on AI agents' long-horizon engineering ability to migrate a complete PyTorch project to JAX and adapt to TPU hardware. Results show current frontier models remain unreliable on this task: Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and all others below 40%.

Product Applications

1. vLLM team explains REINFORCE Zero: achieving zero mismatch between async RL sampling and training

The vLLM official technical blog describes the engineering practice of REINFORCE Zero in eliminating sampling-training mismatch under asynchronous reinforcement learning (Async RL). Building on exact bit-level alignment in synchronous RL, the team further analyzes zero-mismatch mechanisms under weight lag and off-policy settings. The team claims it trains stably for 20,000 steps with policy delay 64, achieving 2.40 times the training throughput of synchronous training, and has covered DeepSeek-V4.1-Flash and Kimi K3.1.

2. GitHub releases agent engineering guide: October weekly commits exceed 1.2 billion

GitHub's official blog released an agent engineering guide for developers, disclosing that in October this year, platform weekly code commits exceeded 1.2 billion, a year-over-year increase of over 400%, exceeding the total commits for all of 2025. The guide notes that the surge in code generation speed has not equally translated into overall R&D productivity, and team workflows need to shift toward higher-level agentic collaboration using Copilot for planning, code review, and online troubleshooting.

3. AWS demonstrates building serverless MCP servers on ECS Fargate

The AWS ML blog presented a method for building serverless Model Context Protocol (MCP) servers on Amazon ECS and AWS Fargate. This approach provides AI agents with secure, metadata-driven access to enterprise data lakes, allowing new datasets to be exposed to the system simply by configuring a JSON metadata file, eliminating the need to write integration code for each dataset.

4. ByteDance's Doubao Work Personal Edition lifts 5-hour usage limit, switches to flexible weekly quota

ByteDance's Doubao Work Personal Edition announced that starting November 6, the original 5-hour short-cycle quota limit is removed. After the adjustment, individual users no longer need to wait for a reset after hitting the short-cycle cap; they can allocate usage within a total weekly quota based on actual business peaks and troughs, supporting workflows such as long-horizon analysis, intensive document writing, and continuous reasoning across stages.

5. Meta launches Spaces for Muse agents, supporting building dedicated interactive task spaces

Meta announced Spaces, a new collaboration feature for its Muse personal AI assistant, allowing users to build highly customized task interaction spaces through natural language. Spaces exists as a sidebar, supports connecting external apps like Gmail, Google Calendar, and Notion, and continuously updates information such as prices and dates in the background.

6. Amazon Alexa launches SMS shopping interaction

Amazon announced SMS interaction for Alexa for Shopping, allowing users to send text or photos to let the AI assistant help with shopping. The system can search Amazon inventory based on descriptions or product photos, and can place orders directly. The feature is launched ahead of the year-end holiday shopping season.

7. DigitalOcean launches Agent Learning and Data Services public previews

DigitalOcean released public previews of Agent Learning Services and Agent Data Services, letting developers build, evaluate, and continuously improve agents on a unified AI-native cloud platform. Learning Services includes Insights for recording agent activity, Signals for identifying failures and performance bottlenecks, and Simulations for testing improvements with simulated conversations before deployment; an enhanced Inference Router optimizes model selection based on task needs.

Industry News

1. OpenAI completes first $11B close of new round at $1.25T valuation

OpenAI has completed the first close of approximately $11 billion in new funding, with a post-money valuation of $1.25 trillion, and the funds have been transferred to the company's accounts. Investors include SoftBank, Sequoia Capital, and Coatue. The total round may increase to up to $20 billion. Due to adjustments in public tech stocks, OpenAI's initial plan to raise $30 billion at a $1.4 trillion valuation was reduced.

2. METR audit: Anthropic's model development speed doubled, possibly triggering its own acceleration threshold

Nonprofit evaluation organization METR conducted an 11-week audit of Anthropic's internal R&D data, concluding that model development speed has roughly doubled from previous levels, with about 80% of the acceleration driven by AI-assisted research, and that Anthropic may have crossed its own AI acceleration warning threshold in July this year. Anthropic obscured the most informative evidence, and there is internal skepticism, so the public must rely more on METR's judgment.

3. OpenAI partners with Hacktron to strengthen frontier system security defenses

OpenAI announced a partnership with security firm Hacktron to use frontier AI models to discover vulnerabilities in its own systems, test connections with third-party services, and harden defenses.

4. Ramp data shows enterprise AI spending nearly stalled in late summer; token spending grew only 1.5% in Aug-Sep

Financial services platform Ramp's latest report shows that enterprise AI spending growth stalled significantly in late summer, with token consumption spending growing just 1.5% between August and September, far below the 42% month-over-month growth recorded in July. The report notes that while token call volumes continue to rise, intense price wars and customers' cost optimization on model calls are compressing revenue growth for major model vendors.

5. Research: free Chinese open-source models now close to closed-source frontier in network attack capability

Capability evaluation company Vals tested open-weight models on multi-step network attack tasks, and results show that open models' attack capability in simulated enterprise networks is approaching the best closed-source frontier models available through trusted access programs. Vals researcher Nikil Ravi said that openly downloadable open-source models are beginning to match lab-internal frontier models.

6. Brokers pitch that AI labs offer up to $2 million for media work records and emails

Wired reports that its reporters have received multiple broker pitches claiming that AI labs such as Anthropic and OpenAI are paying media organizations $350,000 to $2 million per license for proprietary assets like old work documents, internal emails, Slack chats, and meeting minutes. The emails say that with public training data nearly exhausted, labs' spending on private data has exceeded $45 billion.

Open Source Tools

1. Cloudflare and VoidZero launch tsrs: AI-generated Rust version of TypeScript compiler, up to 8x faster than 7.0

tsrs is a Rust port of the TypeScript compiler, developed by Cloudflare, VoidZero, and Max Schwenk from Owner.com, incubated in the Oxc toolchain and in collaboration with the TypeScript team. The port was driven by Claude Opus 5.5 Agent, benchmarked against TypeScript's official 15,197 conformance tests. Cloudflare claims up to 8x faster type checking and ~70% memory reduction relative to TypeScript 7.0, but third-party measurements on 7.1 show about 2.7x.

2. Allen AI open-sources ToyBlocks benchmark: controlled environment for robot learning and evaluation

The Allen Institute for AI (allenai) announced the open-source robot learning research benchmark ToyBlocks, releasing code, datasets, models, and documentation. The project aims to solve the problem of tracing generalization failure causes in traditional robot demonstration tasks, providing researchers with a controlled physical experiment environment with precisely controllable variables, reproducible tests, and traceable model design flaws.

3. Tencent open-sources cloud-native self-evolving AI development platform LoopForge

Tencent's Lightweight Cloud team open-sourced LoopForge, a cloud-native self-evolving AI development platform for real R&D team collaboration. It addresses bottlenecks such as context breaks in single-machine Coding Agents, code going out of control, and repeated pitfalls of similar issues, integrating 7x24 cloud workspaces and multi-role collaboration.

Policy and Opinion

1. MIIT solicits public comments on robot "ID card" measures, plans full-category one-machine-one-code

MIIT's Equipment Industry Division One is soliciting public comments on the "Robot Product Identification Code Management Measures (Draft for Comments)" with the public comment period until December 5. The document proposes a unified robot product identity system, extending the previous coding rules that only applied to humanoid robots to all categories including industrial, service, and special robots. The code consists of four parts: manufacturer, product description, sequence number, and other codes, based on national standard GB/T 48541 to achieve full lifecycle traceability from factory to scrapping.

2. Meta pauses business terms update amid backlash, indefinitely postpones forcing AI training on external links

After widespread criticism, Meta has indefinitely postponed its proposed business terms revision. The original plan would have required Facebook business accounts to authorize Meta to broadly train AI on content from external linked websites. A Meta spokesperson confirmed to Reuters that the company has paused the changes, current terms remain in effect, and it will continue discussions with stakeholders.

---

This article is auto-aggregated by AI225 AI Daily, collecting 27 AI developments. Content is AI-summarized based on public information and may contain omissions or errors, for reference only. Full content and related sources visit: https://daily.ai225.com/daily/2026-11-07