AI Today: DeepSeek V4.1-Flash and Kimi K3.1 Pro
Authorship and approval
Published and updated
State of knowledge as of 2026-11-12. Source material gathered between 2026-11-11 and 2026-11-12. The summary covers two open-weight models, voice assistant deployments, and a video generation tool. DeepSeek reports improved V4.1-Flash scores across 11 benchmarks, while Kimi K3.1 Pro features a one-million-token context window. Meta rolls out Muse conversations in the US, Google brings Gemini Live to CarPlay, and Gemini Omni 1.2 Flash extends generated clips to 15 seconds.
Highlights Today (TL;DR)
- DeepSeek V4.1-Flash-1017: This version achieved higher scores than the previous one across 11 benchmarks.
- Kimi K3.1 Pro: The model handles text and images and features a one-million-token context window.
- Muse: Meta rolls out real-time voice conversations in the app in the US, on phones, in the browser, and on Mac.
- Gemini Live in Apple CarPlay: iOS Gemini users can chat with the assistant hands-free.
- Gemini Omni 1.2 Flash: The model generates clips up to 15 seconds long, with prices dropping by 30% compared to version 1.1.
DeepSeek Updates V4.1-Flash, Improving Scores in 11 Benchmarks
DeepSeek released V4.1-Flash-1017 on Hugging Face, achieving higher scores than the previous version across 11 benchmarks. The report claims faster execution times and shorter responses, allowing end-users to finish tasks faster and at a lower cost.
- Terminal-Bench 4.0: the score increased from 31.2 to 42.3; the test evaluates tasks performed in the terminal.
- ProgramBench: the score increased from 20.3 to 29.1, and Automation-Bench from 54.8 to 60.5.
- Chartography with tools: the score increased from 78.9 to 83.0; BabyVision from 89.6 to 92.5.
- The context window, representing the amount of text processed at once, is up to 1 million tokens. The recommended maximum response length is at least 256k tokens.
- The model weights are under the MIT license, and the repository includes instructions for local execution. The source does not specify memory and hardware requirements. In the Lab: The lab database contains the DeepSeek-V4.1-Flash description, indicating a context of up to one million tokens and KV cache compression. The new report presents version 1017 as an update to this model, featuring improved benchmark scores. For AI Builders: Teams running the model locally can check weight conversion instructions and prompt encoding, which is the format of instructions passed to the model. The source does not provide API pricing, hardware requirements, or availability in the EU.
Moonshot AI Released Kimi K3.1 Pro with 1M Context
Moonshot AI has released the Kimi K3.1 Pro model on Hugging Face with a one-million-token context window. The model supports text and images, and its weights are open.
- 2.8 trillion parameters in the MoE architecture, meaning the model selects a portion of specialized components; 104 billion are active.
- Kimi K3.1 Pro was created by further training Kimi K3 with reinforcement learning across hundreds of tasks. This is the first model in the K3.1 series.
- In benchmarks, it achieved 70 points in HLE-Full with tools, 92.1 in BrowseComp, and 51.7 in SWE Marathon v1.1.
- The source does not specify memory and hardware requirements for running with open weights. In the Lab: The lab database features the September Kimi-K3 deployment comparison: it described the 2.8T model with a 1M context, so K3.1 maintains this scale while introducing a new version and claimed improvements. For AI Builders: Teams considering running the model with their own weights should check the hardware requirements and license. The material also does not provide pricing or results for the Polish language.
Meta Deploys Muse Voice Chats in the US
Meta is rolling out real-time voice conversations to the Muse app on phones, in the browser, and on Mac. Users can speak with the assistant, interrupt it, and choose whether they want to see an avatar.
- Muse App: the feature is being rolled out in the United States.
- Avatar: reacts with facial expressions and gestures matching the tone of the conversation; it can be displayed next to the chat or in full screen.
- Hundreds of voices: the user can choose a voice style and change it at any time.
- AI Glasses and Muse Charm: voice conversations are planned to come to these devices in the future. In the Lab: the lab database contains the Muse agent preview for AI glasses and the Muse Charm voice device; the current deployment brings voice chats to the app. For AI Builders: Teams building voice assistants can check Muse's availability in the US and test interruption-capable conversations. The source does not provide pricing, Polish language support, limits, or EU availability timelines.
Google Brings Gemini Live to Apple CarPlay
Google is introducing Gemini Live to Apple CarPlay so drivers can talk to the assistant hands-free. iPhone users can check their calendar, summarize emails, and practice for meetings using voice commands.
- Gemini Live supports free-flowing conversation: you can change topics, add details, and take breaks.
- Conversations sync with the Gemini app on the iPhone; they wait in your chat history after you park.
- The rollout is starting globally for Gemini users on iOS in supported Gemini Live languages.
- To get started, you need to update the app, enable microphone access, and use Gemini Live on your phone beforehand.
- Daily Brief requires prior enablement. Personalization of this feature is not available for users under 18, nor in the EEA and Nigeria. In the Lab: Google previously described Gemini Live voice productivity features. CarPlay brings a similar workflow with the assistant to the car display. For AI Builders: Teams building voice assistants should verify language support in their target markets. The source does not provide pricing or a full list of languages.
Google Extends Gemini Omni Flash Clips to 15 Seconds
Google has released Gemini Omni 1.2 Flash, which generates 15-second videos at a price 30% lower than version 1.1. Developers and businesses can access the model via the Gemini API, Google Vids, and Gemini for business.
- 3–15 seconds: clip length can be set with single-second precision.
- $0.07 per second in 720p — that is the cost of generation after the price drop.
- 360p or 720p is used for quick sketches; the final clip can be upscaled to 1080p or 4K.
- Conversational editing allows refining clips with follow-up prompts instead of generating them from scratch. You can change the background, lighting, and aspect ratio, among other things.
- The model accepts reference images, short motion clips up to 3 seconds, and audio signals; Google claims better consistency for characters, objects, and scenes. In the Lab: On August 28, we covered Gemini Omni 1.1 Flash lets you build with more control; version 1.2 extends clips from 10 to 15 seconds and lowers the price by 30%. For AI Builders: Teams creating video ads can check how the per-second price translates to production costs and whether follow-up prompt editing fits their workflow. The announcement does not specify usage limits, availability in Poland, or license terms.
Frequently Asked Questions (FAQ)
Which benchmark scores did DeepSeek V4.1-Flash-1017 improve?
The V4.1-Flash-1017 report shows higher scores than the previous version across 11 benchmarks. Terminal-Bench increased from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and Chartography with tools from 78.9 to 83.0. The report also claims shorter execution times and responses.
What is known about Kimi K3.1 Pro?
The model features 2.8 trillion parameters in an MoE architecture with 104 billion active, and a one-million-token context window. It supports text and images. The material provides scores in HLE-Full, BrowseComp, and SWE Marathon, but omits hardware requirements.
How has video generation changed in Gemini Omni 1.2 Flash?
The model creates clips ranging from 3 to 15 seconds, with 720p generation costing $0.07 per second. Version 1.2 is 30% cheaper than 1.1 and is available via the Gemini API, Google Vids, and Gemini for business, among others.
Sources
- DeepSeek: DeepSeek-V4.1-Flash-1017 — 2026-10-17
- Kimi (Moonshot AI): Kimi-K3.1-Pro — 2026-11-02
- Give Your Muse a Voice — 2026-11-05
- Talk it through with Gemini, now on Apple CarPlay — 2026-11-11
- Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips — 2026-11-05 --- Text prepared automatically by the aipoc-lab (AI) laboratory based on publicly available sources listed above. Independent summaries, not copies. Pre-human-editing version.
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
AI Today: ChatGPT Wallet and DeepSeek API
Discover how ChatGPT Wallet, DeepSeek API, Google WeatherNext 3, and Gemini updates shape AI applications in mid-November 2026.
AI Today: Gemini 4 Hydrogen and AlphaGenome 2
Google releases Gemini 4 Hydrogen and AlphaGenome 2, while Meta introduces user-customizable spaces in the Muse app.
AI Today: Gemini 4 Hydrogen and OpenAI Agents
Discover Gemini 4 Hydrogen with a million-token context, OpenAI healthcare agents, NVIDIA's DGX Spark cluster, and Google Ask Maps updates.
AI Today: llama.cpp 1.0.0 and Shield
Discover the latest AI updates including llama.cpp 1.0.0, Shield for IT portals, the XYBench benchmark, Isaac Sim, and Dyania Health.