Today’s through-line is open-weight flagships, sovereign capital in frontier labs, rewritten behavioral rules, and consolidation in voice and inference infrastructure: Mistral opens Large 4 weights, Cohere pushes a CAD $7 billion-scale round with Canadian state equity, OpenAI updates its Model Spec, Deepgram acquires Daily, and FuriosaAI closes a large Series D.

1. Mistral opens Large 4 (Le Chonk) weights: coding and defensive security in one checkpoint

On October 29, 2026, Mistral AI released the weights of Mistral Large 4 (codename Le Chonk). The model is a natively multimodal mixture-of-experts system with roughly 1 trillion total parameters, about 49 billion active parameters per token (about 52 billion including embeddings and output layers), 128 routed experts with four active per token plus one shared expert, a 1-million-token context, text-and-image input with text output, and claimed fluency in 160-plus languages including every official EU language. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters.

Weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B (with an NVFP4 variant), with day-0 support in transformers, vLLM, llama.cpp and TRL. The license is the MIT-derived Mistral Ecosystem Open-Weight License: use, modification, fine-tuning and commercialization are allowed, but operating a Model-as-a-Service business requires a separate commercial license. The API serves mistral-large-4 at $0.70 input / $0.07 cached input / $2.10 output per million tokens.

Company-reported benchmarks include 19.7% on Cognition’s FrontierCode 2.0 (top open-weight, per Mistral), about 48% on the Artificial Analysis Coding Agent Index (effectively tied with GLM-5.3 while using roughly 47% fewer tokens per task), and 41% on Terminal-Bench 4.0 (+13 percentage points versus preview, with about 30% fewer output tokens). On security, Mistral cites 51% on the Artificial Analysis Cyber Index and an average 96.6% refusal rate on malicious cyber prompts across JailbreakBench, StrongREJECT and AgentHarm—highest among open models in its measurement. Three reasoning modes (low / medium / high) are available; a specialized cybersecurity preview variant is offered to vetted design partners such as Orange Cyberdefense.

Takeaway: A trillion-parameter open release ties self-hosted coding agents to auditable defensive security, while the MaaS carve-out keeps a commercial gate on hosted resale.

Sources: Mistral; Hugging Face blog; model card

2. Cohere targets up to CAD $7B Series E; Canada plans a CAD $1.6B stake

Per BetaKit, Toronto-based Cohere on October 29, 2026 announced at Montréal’s Mila an advancing Series E of up to CAD $7 billion at a nearly CAD $36 billion post-money valuation; a spokesperson said most of the target is secured. The Government of Canada plans a CAD $1.6 billion stake on commercial terms via the Canada Development Investment Corporation (CDEV), staged against milestones such as jobs and revenue. Schwarz Group is leading; new participants include the federal government, Bell Ventures and Bullhound Capital; returning backers include Inovia Capital, NVIDIA and Prosperity7 Ventures.

CEO Aidan Gomez said the round would make Cohere “the largest AI company outside of the United States,” keep headquarters in Canada, and support more than 1,000 high-paying tech jobs by 2030; Cohere also pledged CAD $1.7 billion of Canadian compute investment by 2030. Reporting says a merger with Germany’s Aleph Alpha is in final stages (Schwarz is a major Aleph Alpha shareholder), following Cohere’s purchase of Czech cybersecurity AI startup Aisle. If completed at the stated scale, the round would be Canada’s largest startup financing and make Cohere the country’s most valuable private technology firm.

Takeaway: Sovereign capital, European industrial money and a frontier lab are locking equity to a dual story of domestic compute and a transatlantic alternative to U.S. giants.

Source: BetaKit

3. OpenAI updates the Model Spec around five values

On October 29, 2026, OpenAI published an updated Model Spec that reorganizes public behavioral guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and retires most detailed style rules so style can adapt to user preference; throughout, the model is now called “GPT” rather than “the assistant.”

Faithfulness is foundational: OpenAI sets the Spec and boundaries, developers direct applications within them, and users direct work subject to those constraints. The update refreshes the chain of command and stresses careful use of conversation, memory and other context when intent is incomplete. OpenAI also says it is exploring training methods that teach these core values more directly. The Spec describes target behavior; current models do not yet fully reflect it.

Takeaway: Moving from style checklists to a value framework concentrates product defaults and safety explanations in trade-offs among intent, harm and agency.

Sources: OpenAI; Model Spec 2026/10/29

4. Deepgram acquires Daily; Pipecat stays open with a foundation

On October 29, 2026, voice AI infrastructure company Deepgram announced it is acquiring Daily, creator of the open-source voice-agent framework Pipecat; terms were not disclosed. Pipecat will remain fully open source under BSD V2 and vendor-neutral, with an independent Pipecat Foundation whose steering committee can include Deepgram competitors. Daily’s team joins Deepgram; co-founder Kwindla Hultman Kramer becomes Head of Open Source. Deepgram plans a Voice Agent Platform combining its speech models, Pipecat and Daily’s realtime network, and introduced the Real-Time Context Standard (RCS), an open protocol co-developed with Pipecat for carrying conversational context across the stack.

Takeaway: Stack consolidation need not mean framework lock-in—buyers are using foundation and license commitments to blunt “acquihire then close” fears.

Sources: Deepgram; Daily; TechCrunch

5. FuriosaAI closes $620M Series D as RNGD hits volume production

Korea’s FuriosaAI announced a $620 million Series D at a $2.9 billion post-money valuation (company blog dated October 28, 2026). International investors include Infinitum, Jericho Capital, QUDA Capital, Cisco Investments, NeoFort Holding and Worldline Capital; Korea’s National Growth Fund committed about $280 million (roughly KRW 370 billion) as the largest single commitment, with new domestic capital from the Export-Import Bank of Korea, UAMCO and Hanwha Asset Management, and returning backers including NAVER Ventures, DSC Investment and KDB Capital.

Furiosa says its RNGD inference accelerator is in mass production, with about 20,000 cards planned this year and fab/memory capacity to more than double in 2027; the card is a ~180 W air-cooled PCIe part, with eight cards forming the NXT RNGD Server. Proceeds fund global deployments and a third-generation chip with Broadcom and partners—2nm multi-die with HBM4/4E on Furiosa’s Tensor Contraction Processor architecture—aimed at large-cluster agentic inference. Named customers include Samsung SDS (NPU-as-a-Service) and LG Group on-prem hybrid clouds. In the same funding window, cloud-security firm Upwind announced a $400 million Series C at about $3.8 billion, underscoring continued capital intensity in inference silicon and AI security.

Takeaway: Under GPU supply and power constraints, vendors that have moved second-generation inference silicon into customer racks are competing for non-NVIDIA share on efficiency plus programmability.

Sources: FuriosaAI; Business Wire

---

One-line takeaway: Mistral puts trillion-parameter Large 4 weights into self-hosted coding and defense hands, Cohere ties a CAD $7B-scale raise to Canadian sovereign compute, OpenAI rewrites the Model Spec around five values, Deepgram buys Daily while fencing Pipecat open, and FuriosaAI’s $620M Series D pushes RNGD from production into broader deployment.