VIEN

LLM & Agent · Briefing for November 7, 2026

12 stories · 5 sections ·

The post-midterm AI stock selloff is weighing on valuations: OpenAI has reportedly raised only $11 billion so far at a $1.25 trillion valuation, while investors question the timing and price of Anthropic's IPO. On the product side, Thinking Machines launched Inkling 2 Small (594B parameters, open weights coming) and Baidu opened ERNIE 5.5 Flash Base, while Perplexity, Meta, Mistral and Cursor all pushed always-on agents. In security, Google DeepMind used AI agents to find weaknesses in 1024-bit RSA.

Business & Investment

1. OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation

Per Bloomberg (relayed by TechCrunch on Nov 6), OpenAI's round has not closed and could total about $20 billion, short of the $30 billion at $1.4 trillion reported in September. It is the clearest sign yet that the AI stock selloff and data center backlash after the midterms are pulling down top labs' valuations.

TechCrunch

2. Investors doubt Anthropic's IPO timing; Ramp data shows AI spending stalling

Reuters (Nov 6) reports potential investors doubt Anthropic can start its roadshow the week of Nov 9 and list in November; they want a better price and are waiting for October revenue, left blank in the S-1. The same day, Ramp data showed corporate token spending rose only 1.5% across August–September even as token volumes grew, driven by steep price cuts and a shift to cheaper models.

Reuters (via BNN Bloomberg) · TechCrunch

New Models

3. Thinking Machines launches Inkling 2 Small, promises Apache 2.0 weights

Mira Murati's lab announced Inkling 2 Small, a 594B-parameter MoE (17B active) with image input, trained natively in NVFP4, scoring 72.1% on DeepSWE and about 31 on the Artificial Analysis Intelligence Index. It is available via Tinker now, with full weights planned next week after external safety testing; if on schedule, it will be among the strongest open models outside China.

Thinking Machines Lab · Artificial Analysis

4. Baidu releases open weights for 280B-parameter ERNIE 5.5 Flash Base

Baidu put the base (not post-trained) ERNIE 5.5 Flash on Hugging Face: a 280B-parameter MoE with 14B active, pretrained on 33.5 trillion tokens and supporting up to 1M tokens of context in fine-tuning. Releasing the base model lets developers adapt it themselves, competing directly with DeepSeek V4 Flash and MiMo V2.5.

The Decoder

Agents & Tools

5. Perplexity launches Teammates, always-on agents for work

Teammates are personal agents built on Perplexity Computer; each takes its own role, runs in parallel in an isolated environment, can manage and add other teammates, and notifies the user when a decision is needed. It is limited to Max subscribers ($200/month) on web and iOS, competing with OpenAI's and Anthropic's always-on agents.

Engadget · Dataconomy

6. Meta adds Spaces to its Muse agent

Spaces are topic-focused conversations (trip planning, weddings, gift tracking, etc.) where Muse builds trackers, planners or project boards, connects to third-party apps and updates in the background. It arrives a day after Muse Voice; Meta says Muse agents have 10 million users in the U.S. and Canada.

TechCrunch

7. Mistral turns Studio into an enterprise agent platform

Mistral moved Managed Agents (built with the Agents SDK, running durably and resuming after crashes), a natural-language workflow builder, an Apps CLI and tracing into public preview, plus HTTP API connectors that need no MCP server. Private cloud and on-premises deployment are planned for late November, an advantage with European customers that need data sovereignty.

Mistral Release Notes · Mistral AI on X

8. Cursor's cloud agent can now build and deploy live web apps

Build Live Apps lets Cursor's cloud agent scaffold an app from a description, add a Postgres database, put sign-in behind SSO and host it at a shareable link, with an AI security review before publishing. It puts Cursor in direct competition with Lovable and Replit for prompt-to-product.

Cursor Changelog · DEV Community

Security & Safety

9. Google DeepMind uses AI agents to find two weaknesses in 1024-bit RSA

DeepMind researchers with Nadia Heninger (UC San Diego) and Emmanuel Thomé (Inria) say agents, under human supervision, found an algorithm an estimated ~1,600x faster for one short-signature case and cut the cost of batch-factoring 1024-bit keys by about 30x. The results apply only to obsolete 1024-bit RSA, but show AI agents starting to contribute substantively to cryptanalysis.

OfficeChai

10. Hundreds of Artex AI hacking tool deployments found exposed online

Censys found more than 800 deployments of Artex, an open-source AI agent system that can attack from reconnaissance through data exfiltration, reachable without a password. Artex has been linked to attacks on South Korean financial firms and ConnectWise; the project says it will go closed-source, but many copies remain in circulation.

TechCrunch

Vietnam

11. e-Conomy SEA 2026: Vietnam has 45 million generative AI users

The Google, Temasek and Bain report says 45 million Vietnamese (65% of the workforce) use generative AI, 2 in 5 users use it daily, Vietnamese accounts for 80% of AI conversations, and 4 in 5 users are willing to delegate shopping to AI agents. Vietnam's digital economy is expected to reach $49 billion in 2026, up 20%.

VnExpress

12. AIVortex wins AI Arena: Viet Nam 2026

The contest drew 351 teams with nearly 900 students from 80 universities, building proof-of-concept products on Gemini and Google AI Studio for practical problems. The winning team gets a study trip to Singapore including a visit to Google's offices.

Nhan Dan

Latest AI researchAI Research · Paper roundup for November 12, 202610 papers · Thursday, November 12, 2026 →