The Oct 7 AI briefing rounds up the past 24 hours: money and models are both accelerating. DeepSeek's latest funding round is reported to reach at least ¥80 billion, Mistral has put its trillion-parameter flagship into public preview, and Google, OpenAI and Anthropic have each moved on embedding models, mathematics research and office collaboration. All seven items below were published or reported on Oct 6-7, 2026, and can be traced to their original sources.
DeepSeek reportedly raising at least ¥80 billion, with Tencent and CATL in
Bloomberg reported on Oct 6, with Reuters following the same day, that DeepSeek is close to securing at least ¥80 billion (about $11.9 billion), above its original target of roughly ¥50 billion, with the final tally possibly approaching ¥100 billion. Tencent and battery maker CATL are said to be among the largest backers, and the reports also mention preparations for a domestic listing. None of the three companies has publicly confirmed the figures, which come from people familiar with the matter and could still change before closing. Even so, the funding race among leading Chinese model makers has moved up another notch.
Mistral Large 4 enters public preview; weights land later this month
Mistral launched a public preview of its flagship Mistral Large 4, nicknamed Le Chonk, on Oct 6: roughly 1 trillion total parameters with 49B active per token, natively multimodal, available now through the Mistral API and OpenRouter, with open weights planned for late October. On Oct 7, Arena posted its Code Arena WebDev result at 1,534 points, ranked 45th, a large jump over the previous generation. Teams that want to self-host still need to wait for the weights; the preview is best used to test real performance on tasks such as coding.
EmbeddingGemma 2 puts multimodal retrieval on the device
Google DeepMind released EmbeddingGemma 2 on its official blog on Oct 6: built on the Gemma 4 architecture with 740M parameters and released under Apache 2.0, it maps text, code, images, video and audio into one shared embedding space. Quantized, the full multimodal version needs about 567MB of active RAM on a phone, and context grows to 8K tokens. For developers building local search and offline RAG, it is one of the few multimodal embedding options that runs directly on-device.
OpenAI shares mathematics results from an internal frontier model
On Oct 7, OpenAI published a set of new mathematical results produced by an internal frontier model, released through a GitHub repository with protocols for paper revisions and citations; many proofs are formalized in Lean so a computer can verify them. The company also shared reasoning summaries and compute estimates, putting the average result at roughly three hours of ChatGPT Pro thinking. The model itself is not yet released, so the drop reads as a verifiable showcase of research capability.
Anthropic ships two updates: Claude moves into Google Workspace
Anthropic opened the Claude for Google Workspace public beta on Oct 6: one add-on install puts Claude in a sidebar inside Docs, Sheets and Slides, where it can read the open file and edit text, write formulas and build slides directly, asking for approval before changes by default. A Claude Code cloud sessions guide published the same day moves coding tasks to the cloud: each task gets its own virtual machine and branch, several can run in parallel, and work continues with the laptop closed. Pro and Max users also get a limited-time bonus credit, claimable until Oct 7.
ChatGPT gains a Meetings plug-in that follows up for you
An OpenAI team member announced on Oct 6 that ChatGPT's Meetings plug-in is rolling out in beta: it takes meeting notes, then uses what ChatGPT knows about your work to produce personalized summaries and next steps saved to a ChatGPT Space, kept private or shared with the team, and can update project plans or draft follow-ups from them. It is available to Pro and Business users first, with Enterprise to follow. Meeting tools are moving from transcription to actually advancing the next step.
GitHub rebuilds Git infrastructure as agents multiply commits
GitHub's engineering blog disclosed on Oct 6 that monthly Git events hit 473.3 billion in August 2026, more than double the earlier level; September saw 7.38 billion commits and pushes up 4.9x year over year, with concurrent agent reads and writes becoming the new bottleneck. GitHub is rebuilding its Git infrastructure, separating durable storage from compute, with internal benchmarks showing up to 35x higher write throughput. For teams leaning heavily on coding agents, the platform layer is being re-founded for agent-scale development.
