Google’s Gemini 4 Carbon, Mistral’s open-weight Large 4

 

**Google’s Gemini 4 Carbon, Mistral’s 

open-

weight Large 4 (“Le 

Chonk”), Cohere’s large 

sovereign funding round (

with a Canadian 

government stake),

### Model Releases

- **Google announced Gemini 4 Carbon** (Oct 28), ranking #1 on the Artificial Analysis Intelligence Index v5 at 59. It posts strong cybersecurity results (88.2% on SEC-bench Pro for vulnerability finding, 54.8% on ExploitGym, 71% on CWE-bench for patching) and gains on coding/agentic/long-horizon tasks. Introductory pricing is $2 per million input tokens / $10 per million output; initial access is limited to vetted cyber defenders via the Fairwind Program, with broader availability later. This matters because it reclaims the independent overall lead shortly after Gemini 4 Argon and prioritizes defensive cyber capabilities amid rising real-world AI-assisted security risks.

- **Mistral released open weights for Mistral Large 4 (“Le Chonk”)** (Oct 29). It is a ~1-trillion-parameter natively multimodal Mixture-of-Experts model (~49–52B active parameters), trained from scratch on European infrastructure. It shows solid gains in agentic coding (e.g., Terminal-Bench 4.0 improved to ~41%) and defensive cybersecurity while remaining competitive on efficiency. Weights are available under a modified MIT-style license (with a MaaS carve-out). This matters because it advances the open-weight frontier with a large, self-hostable European-trained model, supporting sovereignty and customization without relying solely on closed APIs or Chinese bases.

- **Nvidia released Nemotron 3.5 Transcribe** (Oct 28), an open-weight ~30B-parameter MoE speech model (~4.3B active). It handles transcription in 40 languages, speaker diarization (up to 8 speakers), translation (31 languages to/from English), and audio Q&A on recordings up to ~1 hour, controllable via natural-language prompts. It leads or ranks highly among open models on relevant ASR and diarization leaderboards and runs on a single GPU. This matters because it consolidates multiple speech tasks into one efficient, open, promptable model suitable for local or scaled use.

### Funding Rounds

- **Cohere announced the initial close of a Series E** targeting up to CAD $7 billion (~$5 billion USD), implying a valuation around CAD $35.7 billion. Participants include Schwarz Group (lead), Nvidia, Bell Canada, and the Canadian government (equity investment of up to $1.6 billion). The round supports models, regulated-industry products, and Canadian compute infrastructure; Cohere commits to remaining headquartered in Canada, creating jobs, and investing in local infrastructure. This matters because it is one of the largest Canadian tech raises and represents a significant government equity stake to build sovereign AI capacity outside the U.S. and China.

### Policy Changes / Regulatory

- **The U.S. Department of Justice is inquiring about Nvidia’s planned ~$12.9–13 billion acquisition of Hugging Face** (reported Oct 28), including whether it would let Nvidia control a key distribution hub for open AI models. This sits within a broader, ongoing probe of Nvidia’s conduct (acquisitions, sales practices, customer pressure) that began under the prior administration. Nvidia expects the deal to close in H1 2027 and has said Hugging Face will remain multi-provider. This matters because Hugging Face is a central open-model platform; regulatory pushback could affect the open ecosystem and Nvidia’s expanding vertical integration in AI infrastructure.

No major standalone research-paper breakthroughs dominated the last 24 hours in public reporting (AI-assisted mathematical results continue to appear but are not the headline). Minor product updates, unconfirmed distillation claims, and older funding items (e.g., Meta’s large equity raise from earlier in the week) were excluded. Sources include official announcements and contemporaneous reporting from Reuters, company blogs, and tech outlets.