2026-06-16 · TuesdayPublished daily at 08:00
AI Daily
- 1MiniMax Releases Open-Source M3 Model Weights and MSA Technical PaperOfficial account: MiniMax
- 2
- 3
01Model Releases/Updates
MiniMax Releases Open-Source M3 Model Weights and MSA Technical Paper
Last Friday, MiniMax open-sourced the weights for M3, which has 428B total parameters and 23B active parameters, alongside a technical paper on MSA (MiniMax Sparse Attention), an architecture that significantly reduces long-context computation costs. M3 is the first open-source model trained with interleaved multimodal mixtures of text, images, and other modalities from the pretraining stage. Two weeks after release, M3 ranked first among open-source models on both the Artificial Analysis Intelligence Index and the GDPval-AA leaderboard, reached the Pareto frontier on Code Arena WebDev, and ranked first among Chinese models on the Vals.AI leaderboard. Output speed has increased from about 30 TPS to about 80 TPS, with a further 30–40% increase planned; a usage dashboard has been added to the Token Plan console.
Original textNext-Generation Speculative Decoding: DFlash and Spec V2
Z Lab, Modal, and the SGLang team jointly released the DFlash speculative decoding model and SGLang's default Spec V2 engine. DFlash uses block diffusion + KV injection to generate entire blocks of draft tokens in parallel. On the HumanEval dataset with Qwen 3.5 397B-A17B (BF16) at concurrency 1, throughput reaches 4.3 times the baseline
Original text02Product Releases/Updates
Flash-KMeans: IO-Aware Exact K-Means, Over 200 Times Faster Than FAISS on GPUs
Teams from UC Berkeley and UT Austin have open-sourced Flash-KMeans (Apache 2.0, `pip install flash-kmeans`), an exact implementation of standard Lloyd's k-Means that accelerates computation by restructuring GPU data flow rather than changing the mathematics or using approximations. On NVIDIA H200, end-to-end performance is 17.9× faster than the best baseline, 33× faster than cuML, and over 200× faster than FAISS. Its FlashAssign kernel avoids materializing the full N×K distance matrix, reducing IO complexity from O(NK) to O(Nd+Kd) and delivering up to 21.2× speedup for the kernel; the Sort-Inverse Update kernel reduces atomic contention by sorting cluster IDs, delivering up to 6.3× speedup for the kernel. It supports out-of-core processing, with a single iteration on 1B data points at K=32768…
Original textGrok Build Introduces Agent Dashboard to Manage Multiple Coding Sessions
xAI has introduced Agent Dashboard for Grok Build, providing a single screen for managing multiple coding sessions. The dashboard groups sessions by status (awaiting input, working, idle), with each row showing a status indicator, name, branch, permission mode, and current action. Selecting a session opens a peek panel to view its latest output and reply directly. For sessions awaiting input, options can be selected using arrow keys or number keys. The input box at the bottom lets you assign new sessions, with options to set the model, start plan mode, or automatically approve edits. Open it with `grok dashboard`, `/dashboard`, or `Ctrl+\`. Sessions continue running after the dashboard is closed, and reopening it restores the view.
Original textMeta Launches “AI Mode” on Facebook to Synthesize Answers from Public Information on the Platform
Meta announced the launch of the “AI Mode” search feature on Facebook, using Meta AI to extract information from public posts, including groups and Reels, and synthesize answers. Users can ask questions in natural language to receive summaries. New features also include video collage editing, transition effects, and AI photo presets that can change outfits, hairstyles, and accessories. Sports fans can tap “AI Edit” in Stories to virtually put on team jerseys. These updates continue the deployment pace of previously launched features such as dynamic avatars, automated Marketplace replies, and AI assistants for creators. In addition, Meta recently launched global subscription plans for Facebook, Instagram, and WhatsApp, starting at $3.99 per month, with more AI subscription tiers in development.
Original text6x Faster! Kimi K2.7 Code High-Speed Version Is Now Available
The Kimi K2.7 Code High-Speed Version is now available. It uses the same model as the regular version, with output speeds approximately 5–6 times faster: about 180 Token/s for standard programming scenarios and up to 260 Token/s for short contexts. API pricing is twice that of the regular version. Model ID: kimi-k2.7-code-highspeed. Kimi Code Plan users can access it through the “Early Access Program,” with usage consumption three times that of the regular version. Thinking mode must be enabled; disabling it will cause an error or fall back to K2.6. To celebrate the release, the Kimi API Open Platform is offering a three-week top-up promotion: deposits of 500 yuan or more receive 20%–30% vouchers. Compared with K2.6, K2.7 Code offers improved long-context programming instruction following and long-horizon task performance, reduces average token consumption by 30%, and shows significant gains in internal benchmark tests. Truncated…
Original textOpenRouter Adds Free Models gpt-oss-20b and Gemma4 26B
New free capacity is now available on OpenRouter, provided by Darkbloom from @eigenlabs: gpt-oss-20b and Gemma 4 26B. Start using these models today ↓
Original textGitHub Releases New Open-Source Dataset to Accelerate Multilingual AI Research and Development
GitHub has released a repository-level dataset under the CC0-1.0 license, covering multilingual developer content, including READMEs, issues, and pull requests. The dataset is intended to help researchers and developers discover and leverage technical documentation and community discussions across languages, advancing the development and optimization of multilingual AI.
Original textClaude Code v2.1.178 Released
This update adds the `Tool(param:value)` syntax for permission rules to match tool input parameters; skills in nested `skills` directories are loaded automatically, with name conflicts preserved in the `<dir>:<name>` format; and conflicts among nested `agent`, `workflow`, and `output-style` entries are resolved in favor of the nearest directory. It improves classifier evaluation before sub-agent generation in auto mode; `/doctor` now uses a flat tree layout; workflow prompts are highlighted with a purple flash and trigger only on explicit phrases; and `/bug` now requires a description before submission. It fixes crashes caused by the CLI inheriting stale WebSocket/OAuth file descriptors, connection failures caused by mismatched OAuth token accounts in Chrome, tool results appearing in sub-agent transcripts, and background recovery restarting from the beginning…
Original text03Industry News
Salesforce Acquires AI Customer Service Platform Fin for $3.6 Billion
Salesforce announced the acquisition of AI customer service platform Fin, formerly Intercom, for $3.6 billion. Fin provides AI agents that can solve customer issues across multiple channels, including live chat, WhatsApp, SMS, phone, and Slack. Salesforce plans to use Fin’s technology and team to enhance its enterprise Agentforce platform, which allows businesses to build custom AI agents to automate tasks. The transaction is expected to close in Salesforce’s fourth fiscal quarter of 2027, meaning early 2027. Fin co-founder and CEO Eoghan McCabe will continue as CEO, while head of engineering Des will continue to lead engineering.
Original textNvidia Joins the AI Debt Boom, Issuing $20 Billion in Bonds
Nvidia plans to raise at least $20 billion through its first bond offering since 2021, according to people familiar with the matter. The move marks Nvidia’s entry into the wave of debt financing in the AI sector.
Original textCloudflare Brings in the Ensemble AI Team to Accelerate AI Infrastructure Development
Cloudflare announced that key members of the Ensemble AI team are joining the company to accelerate AI infrastructure development. Ensemble focuses on model compression and efficient inference. It developed NdLinear, which can directly replace standard Transformer linear layers while preserving multidimensional activation structures, and NdLinear-LoRA, which reduces the number of trainable parameters needed to fine-tune large models. These technologies complement methods such as quantization and aim to reduce the memory, compute, and deployment costs of large language models and multimodal architectures. Cloudflare will integrate Ensemble’s work into the Workers AI platform, further improving inference efficiency, GPU utilization, and deployment economics through its global network and serverless GPU inference services.
Original textxAI Announces Grok Integration with the Warp Terminal Development Environment
xAI announced an integration with Warp, an agent development environment built around the terminal and used by nearly a million developers. Users can access Grok models in Warp with a Grok or X Premium subscription, including the `grok-build-0.1` model that powers Grok Build CLI. To set it up: download Warp, connect your SuperGrok subscription in the Agent settings page, and switch to the `grok-build-0.1` model. More agents and integrations are coming soon.
Original text04Tips and perspectives
Skydio CEO Adam Bry: Silicon Valley Shouldn’t Draw Red Lines Around Drone Use
Skydio is the largest drone manufacturer in the U.S., focused on enterprise markets such as public safety, the military, energy, and infrastructure inspection. CEO Adam Bry said that after the Trump administration banned drones made in China late last year, inexpensive consumer drones nearly disappeared, making Skydio products a major alternative. The company believes drones are shifting from tools to autonomous infrastructure, with large-scale use enabled by docks, remote operation, and software integration, and AI playing a key role. The interview also covered Skydio’s approach to working with the military and how autonomous technology is driving the company’s expansion.
Original textThe Wave of AI Layoffs Becomes a Powder Keg
Tech companies have cut approximately 150,000 jobs so far this year, averaging 974 per day—a pace 44% faster than last year. Nearly 40,000 layoffs last month marked a two-year high, with AI cited as the leading reason for layoffs for three consecutive months. After Block laid off nearly half its workforce, CEO Jack Dorsey denied that AI was the root cause, while Marc Andreessen called AI merely a “silver-bullet excuse.” Uber cut 23% of its HR department, although its CTO had previously revealed that its AI coding budget was exhausted within four months. Meanwhile, AI chipmaker Cerebras reached a market capitalization of $67 billion on its first trading day, SpaceX debuted at a market capitalization of $2.1 trillion, and Anthropic and OpenAI were each valued at approximately $1 trillion. Meta announced 8,000 layoffs after Zuckerberg purchased a $170 million mansion. Polls show that 65% of voters believe a middle-class life is out of reach, while 76% of Americans rank the cost of living as their top economic concern.
Original textThe Golden Age of AI Applications Is Here: The Fable Ban, Nadella’s Moat Thesis, and Salesforce’s Acquisition of Fin
The U.S. government shut down access to Anthropic’s Fable, making open-source and local models essential; Satya Nadella argues that the AI ecosystem’s moat should be human expertise and the systems surrounding models; Salesforce acquired Fin (formerly Intercom) for $3.6 billion, with Fin using open-source models to deliver cost efficiency. These three developments signal the arrival of a golden age for AI applications. The challenges of building AI applications include choosing among models with different characteristics, such as Kimi K2.6, Qwen 3.6 27b, and GLM 5.1; designing hill-climbing loops for agent systems; and continuously evaluating the performance of the model and loop together to maximize intelligence within a token budget. Companies that master these three skills will dominate this era.
Original textA Beginner’s Guide to GitHub Copilot CLI: An Overview of Common Slash Commands
GitHub Copilot CLI provides beginners with an overview of common slash commands, helping users control the AI agent in their terminal through commands.
Original text