Apple is reportedly set to license a specialized version of Google’s Gemini to handle complex Siri requests, with the model running on Apple’s own servers under its Private Cloud Compute framework.

Routine voice tasks like timers, notes, or messages would remain on device, while multi-app queries—such as “plan a trip with my calendar, flights, and hotel confirmations”—would route to the Gemini tier for server-side reasoning.

Reporting pegs the arrangement at near $1 billion per year and emphasizes that Apple will host the model on its infrastructure rather than Google’s. Apple’s PCC design was introduced in 2024 with a stateless, ephemeral processing approach that aims to avoid retaining personal data after requests are handled.

The goal is a smarter, context-aware assistant that behaves like a true digital helper rather than a command interface. With Gemini in the loop, Siri could summarize documents, draft messages, and coordinate actions across apps using conversational context instead of fixed triggers.

The partnership can be described as a pragmatic way to gain reasoning and long-context capabilities quickly while Apple continues its own model work. Observers also note that Gemini’s multimodal strengths, spanning text, audio, and vision, could support future integrations across devices, including camera workflows and spatial platforms.

Mark Gurman of Bloomberg has reported a target window around spring 2026 for the revamped Siri rollout, with timing aligned to broader Apple Intelligence updates.

Under Private Cloud Compute, Apple describes running advanced models in Apple operated data centers with temporary, isolated compute silos, code transparency measures, and no long term retention of user data after fulfillment. If Apple demonstrates that Gemini operates entirely inside those PCC constraints, it can argue for frontier level performance without handing user data to Google.

Paying a rival for core AI capability is a shift, but it prioritizes experience and integration over owning every layer of the stack. Apple controls the device, identity, and privacy posture, and rents the “brains” to close the quality gap now.

If a Gemini powered Siri ships on Apple’s infrastructure with clear privacy proofs and tangible day to day utility, Apple can claim it preserved its brand promise while catching up to the frontier.

Moonshot AI’s Kimi K2 Thinking represents a significant leap for open-source AI, achieving competitive performance with leading closed systems. Built as a 1 trillion-parameter mixture of experts with 32 billion active parameters per step and a 256K context window, it was trained for interleaved thinking — reasoning in steps while making extensive tool calls.

Released under a modified MIT license, the model is optimized for both efficiency and transparency, making it one of the most accessible high-performance reasoning models to date.

Users report it maintains the tone and style from the earlier K2 Instruct model while adding sophisticated reasoning capabilities. In benchmarks:

The model demonstrates 200–300 sequential tool calls without human intervention. Moonshot used quantization-aware training to enable native INT4 inference, achieving 2x speed improvement while maintaining accuracy. It also supports function calling, JSON mode, and structured output generation.

Available through:

Early testers report successful use for code debugging, complex planning, and policy analysis, with the model explaining its reasoning process step-by-step. It has also been tested for long-form summarization, legal document review, and multi-turn agent workflows.

K2 Thinking demonstrates China's AI labs are closing gaps with Western incumbents faster than anticipated. The combination of open-source release, competitive performance, and efficient architecture positions Moonshot as a significant player in global AI development.

🎙️ Meta’s omni-lingual speech AI handles 1,600+ languages natively

🎨 Artist places an AI-generated print-out into a museum

🧑‍⚖️ ChatGPT facing more lawsuits for its role in self-harm

💰 Softbank bets on itself and sells $5.83B of Nvidia stock

🍃 Yann LeCunn to leave Meta and start his own AI lab

🫰OpenAI seemingly asks for government bailout

Here’s what to do:

✍️ Pick two tasks: a multi-step research question and a structured email summary.

📂 Open both models (Kimi and your favorite) in OpenRouter side by side or open your favorite LLM the normal way and go to kimi.com to test Kimi K2 on the app.

🧭 Run the same prompt twice and note how each handles reasoning and format.

🧑‍⚖️ Score accuracy, control, and speed from 0 to 2 each.

✂️ If tied, ask both for a one-sentence self-critique and rerun once.

🏆 Save the winner and prompt for next week’s workflow.

Click below ⬇️

Did Apple make the right move by partnering with Gemini, and has Kimi K2 finally broken western dominance in AI systems? Let us know your pick and why. See you next time! 🚀

Stay informed, stay curious, and stay ahead with Jumble!

Zoe from Jumble