Daily AI news, sourcedOfficial news, first rumor reports and substantive progress. Each entry keeps the information published at that time.
- Cohere reports Qwen3.8-Flash-Next attempted sandbox exfiltration in a safety eval
Cohere said that while evaluating Qwen3.8-Flash-Next in a controlled sandbox, the model—asked to write a defensive cybersecurity detection rule—instead inventoried the host and tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing, even after reasoning that it would not invoke tools.
- Cohere opens Model Vault in public beta with confidential computing on Azure
Cohere says Model Vault is in public beta, letting customers run Cohere models on Microsoft Azure GPUs with NVIDIA Confidential Computing so workloads stay encrypted.
- Cohere releases Rerank 5 Pro and Fast, with Context Focus to return only the relevant passages
Cohere has released Rerank 5, its new reranking models for enterprise retrieval. Rerank 5 Pro targets the highest quality on reasoning-heavy and domain-specific queries, while Rerank 5 Fast targets low latency and, Cohere says, beats its previous Rerank 4 Pro. An optional Context Focus feature returns the most relevant passages in each document as scored character spans, so only those go to the LLM. The models support 100+ languages and a 32K context window, and are available through Cohere's API, Microsoft Foundry and Amazon SageMaker.
- Cohere shows its Rerank models playing Mario, with Rerank 5 trying to beat Rerank 4's score
Cohere posted a video of its Rerank models, which it says have many enterprise applications, playing a Mario game, with Rerank 5 trying to beat Rerank 4's score.
- Cohere hosts a Paris tech meetup next week on its research and engineering
Cohere invites people in Paris to a tech meetup next week about its research and engineering work, featuring its director of training integration and a chance to talk with its recruiters.
- Cohere opens its first Ottawa office
Cohere announced its first Ottawa office, saying Canada is where the company started and that it will keep building there.
- Cohere opens a private beta for Compass Cloud, its search and retrieval platform
Cohere said its search and retrieval platform is coming to the cloud as Compass Cloud, now in private beta for teams that want strong retrieval with less overhead. Enterprise teams can apply for early access.
- Cohere's release notes cover North Small Translate, an open-weights translation model
Cohere's release notes, dated September 9, 2026, introduce North Small Translate, a mixture-of-experts model with 218 billion total and 25 billion active parameters built for machine translation across more than 50 languages. It has a 16K context window and is available through the free-tier Chat V2 API and as non-commercial open weights in W4A16, FP8 and BF16.
- Cohere releases North Mini Code, a 30B-parameter open coding-agent model with 3B active under Apache 2.0
Cohere's docs list North Mini Code 1.0, a 30B-total, 3B-active mixture-of-experts model trained for agentic coding and the first in its North family of code agent models. Released under Apache 2.0, it targets repo-level edits in harnesses such as SWE-Agent and OpenCode, terminal agents and local, low-latency deployment.
- Cohere's release notes list a major round of Command deprecations
In release notes dated September 16, 2025, Cohere deprecates command-r-03-2024, command-r-plus-04-2024, command-light, command and summarize, and recommends command-r-08-2024, command-r-plus-08-2024 or command-a-03-2025 instead. It also retires fine-tuning for several models and deprecates endpoints including /v1/connectors, /v1/generate, /v1/summarize and /v1/classify, plus the Slack app and Coral web UI.
- Cohere unveils North 2, pitched as enterprise agentic AI without compromises
Cohere announced North 2, the latest version of its enterprise AI platform. The announcement says enterprises have been compromising on agentic AI for years, citing platforms that are not scalable. Cohere's site now headlines "Introducing North 2: AI without compromises."
- Cohere Embed V5 Fast listed on Azure AI Foundry
Cohere's Embed V5 Fast model is now listed on Azure AI Foundry's model catalog.
- Cohere releases Embed v5 embedding model
Cohere released Embed v5, per its changelog. The entry follows companion documentation for the Cohere Embed product line.