Claude's Voice Mode Now Runs on Opus/Sonnet — Email Can Be Sent by Voice
Anthropic has announced that Claude's voice mode has moved to the more powerful Opus and Sonnet models. The new voice mode can access Gmail, Google Calendar, and Slack, letting users compose and send email using only their voice.
Claude is reportedly the only AI that can currently send email directly through a voice assistant. While OpenAI's and Google's voice output still sounds more natural, Claude has pulled a step ahead on practicality. As voice-driven hands-free operation becomes part of everyday life, covering the high-frequency task of composing and sending email is a meaningful milestone.
Hugging Face Releases The Stack v3 — a 114TB Open Code Dataset
Hugging Face has released "The Stack v3," one of the largest openly available code datasets to date, with a total size of approximately 114TB.
The dataset is offered in two forms. One is "stack-v3-train," which has been deduplicated, quality-filtered, and PII-masked, and can be loaded directly with `load_dataset`. The other is "stack-v3-full," which retains all the data for users who want to run their own deduplication and filtering.
Securing high-quality code data remains a key challenge in training open source models, and this dataset looks set to lift the whole community. That said, licenses and copyright handling still require continued care.
SANA-Video 2.0 — Fast Video Generation with Hybrid Linear Attention
SANA-Video 2.0, a video diffusion transformer with 5B and 14B parameter variants, has been announced. Its standout feature is the adoption of "Hybrid Linear-Softmax Attention," which combines linear attention (O(N) computational complexity) with softmax attention at a 3:1 ratio.
This keeps quality on par with a full-softmax model while preserving favorable compute scaling for longer videos. On a single H100, it can generate a 720p, 5-second video in 13.06 seconds — claimed to be 120x faster than Wan 2.2-A14B.
It is also worth noting that the model was designed and trained completely from scratch, rather than being a linearized version of an existing model. It stands as an example of an architecture-level approach delivering results in the push for more efficient video generation AI.
SupraLabs Releases a 5M-Sample Reasoning Corpus — Built for Reasoning Training of Small Models
SupraLabs has released "reasoning-corpus-4K-5M-v1," a dataset aimed at improving the reasoning abilities of small language models (SLMs). It contains 5 million samples, each consisting of four fields — the user prompt, the model's thinking trace (chain-of-thought), the final answer, and the ChatML format.
Every sample is kept within a sequence length of 5K tokens, making the dataset easy to handle for SFT/fine-tuning. It has already recorded more than 1,000 downloads.
Efforts to bring reasoning capabilities to small models are on the rise, and releasing practical training data like this is significant. However, the license terms of the source model still demand attention — it is essential to check any restrictions that apply to distillation.
Malice Lurking in AI-Generated Pull Requests — Attack Patterns Spanning Multiple PRs
A Codacy blog post has highlighted a tactic in which AI-generated pull requests (PRs) hide malicious intent by scattering it across multiple PRs. Each individual PR looks harmless, but combined they constitute malicious behavior — an attack vector that is difficult to detect through conventional code review.
As AI-generated code contributions increase, so does the need for governance frameworks that analyze entire groups of PRs rather than reviewing each PR in isolation. For open source projects in particular, verifying contributor identities and monitoring the combined effects of multiple contributions are likely to become key challenges.
Fleet Engineering — Designing and Supervising a "Fleet" of AI Agents
In Japan's developer community, attention is focusing on the concept of "fleet engineering," which refers to the shift from the stage of conversing with a single coding agent to the stage of running multiple agents simultaneously and designing and supervising that fleet.
The concept has surfaced under several names between 2025 and 2026, and it requires a skill set different from traditional single-agent usage — coordinating multiple agents, dividing tasks, and designing oversight mechanisms. Approaches such as defining roles (role/subagent) in tools like Cursor to stabilize context and quality can also be positioned as part of this trend.
While adding more agents improves productivity, new challenges are also emerging, such as increasingly complex orchestration and cost management.