Maestral, the open-source Mac Dropbox client known for its simple and lightweight syncing, has been retired. Developer Sam Schott archived the project in July 2026, citing lack of time and no longer using Dropbox himself. The current version will continue working for now but will not receive further updates or maintenance.
Maestral, the open-source Mac Dropbox client known for its simple and lightweight syncing, has been retired. Developer Sam Schott archived the project in July 2026, citing lack of time and no longer using Dropbox himself. The current version will continue working for now but will not receive further updates or maintenance.
The article discusses how Dropbox's Mac client has become a bloated application over time, contrasting it with what it once was. It highlights the shift from a simple, lightweight syncing tool to a more complex and resource-heavy program, which many users find excessive for their needs.
Using a Family Feud format, the post argues that big companies like Anthropic, Adobe, and Google fail to build great native Mac apps despite vast resources, often resorting to Electron apps. Only Apple consistently delivers "Mac-assed" software, suggesting an inverse relationship between company size and native Mac app quality.
A look at early Mac app icons shows Apple's convention of tilted rectangles and a hand icon. The author traces how this faded as users adapted, with apps like TeachText and Font/DA Mover breaking the mold, arguing that mandating design rules differs from organic adoption.
This analysis was generated by AI and may contain inaccuracies. Always verify with original sources.
No related papers found.
On July 6, 2026, Maestral, the open-source "splendidly simple" Dropbox client for macOS, was officially retired[^1]. The news was first reported by Simon Willison, citing the project's own announcement. Maestral had been a popular alternative for users seeking a lightweight, native Mac client for Dropbox without the overhead of the official app. Its retirement marks the end of a widely appreciated community-driven project.
<p><strong><a href="https://huggingface.co/tencent/Hy3">tencent/Hy3</a></strong></p> New Apache 2.0 licensed model from Tencent in China:</p> <blockquote> <p>Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. Today, we introduce Hy3, which outperforms similar-siz
Tencent Hunyuan 3 (295B MoE) on 2x NVIDIA DGX Spark: NVFP4 W4A16 + native MTP speculative decoding. First published MTP-on-GB10 numbers, full recipe, KV math, and every bring-up bug documented.
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
No Wikipedia article found.
On July 6, 2026, Tencent released Hy3, a large Mixture-of-Experts (MoE) language model with 295 billion total parameters, of which 21 billion are active per inference step, plus an additional 3.8 billion parameters in a Multi-Token Prediction (MTP) layer. The model is licensed under Apache 2.0 and is available on Hugging Face at tencent/Hy3. The full-precision checkpoint consumes 598 GB of storage, while an FP8 quantized variant weighs approximately 300 GB. Hy3 supports a context length of 256,000 tokens.
Hy3 follows a preview release ("Hy3 Preview") from late April 2026. According to the Tencent Hy Team's announcement, feedback from over 50 products informed an enlarged post-training phase with higher-quality data. The team claims Hy3 "outperforms similar-size models and rivals flagship open-source models with 2–5× parameters". The model is also accessible for free via OpenRouter until July 21, 2026.
The release was covered by Simon Willison, who tested Hy3 on OpenRouter by prompting it to "Generate an SVG of a pelican riding a bicycle." The model produced a flat-style cartoon illustration of a white pelican with a large orange beak riding a red bicycle across a pale blue background. Willison tagged the post with "ai," "generative-ai," "llms," "pelican-riding-a-bicycle," "llm-release," and "ai-in-china".
A secondary GitHub repository, tonyd2wild/Hy3-295B-NVFP4-MTP-2x-DGX-Spark, documents a community effort to run Hy3 on two NVIDIA DGX Spark systems using NVFP4 quantization (W4A16) with native MTP speculative decoding. That repository has 10 stars and identifies itself as "Tencent Hunyuan 3".
Social media platforms—Twitter, Reddit, Weibo, and Zhihu—were queried for mentions of Hy3, Tencent, Mixture-of-Experts, OpenRouter, and related keywords. No posts, quotes, or sentiment data were returned from any of these platforms; all four queries failed to retrieve content. As a result, no meaningful social-media reception analysis can be performed for this release.
A search of academic databases and preprint servers using the keywords "Hy3," "Mixture-of-Experts," "MoE," "Tencent," "295B parameters," and "FP8 quantization" returned zero papers. No arXiv preprints or peer-reviewed publications matching the model name or the specific architectural details have been identified as of the briefing date.
The primary source for the Hy3 announcement is the Hugging Face model page at https://huggingface.co/tencent/Hy3, as cited by Simon Willison on his blog. Willison's post, titled "Maestral, the Open Source Splendidly Simple Mac Dropbox Client, Has Been Retired," was published on July 6, 2026 at 23:57:35 UTC. The post's title refers to the retirement of a separate open-source project (Maestral), but the primary content is the Hy3 model announcement. The URL for the blog entry is https://simonwillison.net/2026/Jul/6/hy3/#atom-everything.
The origin narrative constructed by the research system states: "On July 6, 2026, Maestral, the open-source 'splendidly simple' Dropbox client for macOS, was officially retired. The news was first reported by Simon Willison, citing the project's own announcement. Maestral had been a popular alternative for users seeking a lightweight, native Mac client for Dropbox without the overhead of the official app. Its retirement marks the end of a widely appreciated community-driven project." However, this narrative conflates two separate stories: the Maestral retirement appears to be unrelated to the Hy3 model release. The blog post's headline refers to Maestral, but the majority of the content describes Hy3. It remains unclear whether Willison intended a single combined post or whether two distinct items were aggregated under one headline.
The earliest known publication time for the Hy3 announcement is July 6, 2026.
Tencent is the developer of Hy3, and the company is headquartered in China. The model is hosted on Hugging Face under the organization namespace tencent. The Hugging Face page describes Tencent as being "on a journey to advance and democratize artificial intelligence through open source and open science".
Hy3 is described as a 295B-parameter MoE model with 21B active parameters and a 3.8B MTP layer. MoE architectures activate only a subset of parameters per token, which can reduce computational cost relative to dense models of comparable total size. The MTP (Multi-Token Prediction) layer is a technique that enables the model to predict multiple future tokens simultaneously, which can improve inference speed through speculative decoding.
The full model checkpoint is 598 GB, and the FP8 quantized version is approximately 300 GB. The model supports a 256K-token context window.
Hy3 is available for free on OpenRouter until July 21, 2026. OpenRouter is a platform that provides API access to various large language models.
A community repository (tonyd2wild/Hy3-295B-NVFP4-MTP-2x-DGX-Spark) claims to be the "first published MTP-on-GB10 numbers" and documents a setup running Hy3 on two NVIDIA DGX Spark systems using NVFP4 quantization (W4A16) with native MTP speculative decoding. The repository is written in Shell and has 10 stars on GitHub. Its description explicitly refers to "Tencent Hunyuan 3," suggesting Hy3 may be related to or rebranded from Tencent's earlier Hunyuan series.
No funding information is available for the Hy3 project.
The release of Hy3 represents a significant open-source contribution from Tencent, placing a 295B-parameter MoE model under the permissive Apache 2.0 license. The key architectural highlights are the high ratio of total to active parameters (roughly 14:1), the inclusion of an MTP layer for speculative decoding, and a 256K-token context window. The claim that Hy3 "rivals flagship open-source models with 2–5× parameters" positions it as a compute-efficient alternative to much larger dense models, though independent benchmarks or third-party evaluations are not yet available in the retrieved data.
Several caveats limit the depth of analysis possible at this time. First, no academic papers or technical reports accompany the release, making it impossible to verify performance claims against standard benchmarks (e.g., MMLU, HumanEval, GSM8K). Second, social media platforms returned zero data points, so public reception—especially within the Chinese AI community on Weibo and Zhihu—cannot be assessed. Third, the origin narrative in the research system incorrectly merges the Hy3 model announcement with the unrelated retirement of the Maestral Dropbox client, which may cause confusion. The blog post's headline references Maestral, but the body is dominated by Hy3 content; it is possible that Willison published a combined feed item or that the summarization pipeline conflated two separate posts.
The existence of a community GitHub repository attempting to run Hy3 on consumer-grade DGX Spark hardware suggests early interest from the open-source hardware community. The reference to "Tencent Hunyuan 3" in that repository's description hints that Hy3 may be a continuation or variant of Tencent's Hunyuan line, but no official documentation confirms this.
From a competitive landscape perspective, Hy3 enters a crowded field of large open-weight MoE models, including Mixtral 8×22B (141B total, 39B active), Qwen2.5-72B (dense), and DeepSeek-V2 (236B total, 21B active, MoE). Hy3's 21B active parameters place it in a similar compute class to DeepSeek-V2, while its 295B total parameters suggest a larger expert pool than many contemporaries. The Apache 2.0 license is more permissive than the custom licenses used by some Chinese AI labs (e.g., Qwen's research-only license for certain model sizes), which may facilitate broader adoption.
The free access period on OpenRouter through July 21, 2026 provides a limited window for developers and researchers to evaluate Hy3 without incurring API costs. After that date, pricing and availability are unknown.
In summary, Hy3 is a technically ambitious open-source release from a major Chinese technology firm, with clear architectural innovations in MoE scaling, FP8 quantization, and speculative decoding. However, the absence of academic validation, social-media engagement data, and a standalone technical report means that any assessment of its actual performance relative to competitors rests entirely on Tencent's uncorroborated claims. Independent benchmarking and community feedback will be essential to establish Hy3's position in the rapidly evolving landscape of large language models.
Apple's iOS 27 developer beta 3 enables new "Pace" and "Expressivity" sliders for customizing Siri's AI voices, previously marked as "Coming soon." The article notes that after a month of testing, the default settings prove well-calibrated. Two new voices are available: one female and one male.
The article argues that while Big Tech has merged with the Trump administration and global antitrust enforcement has weakened, U.S. state attorneys general can use federal antitrust laws to take on tech giants. It suggests that states should collaborate with international enforcers, who share the same legal frameworks, to hold companies like Meta and Google accountable.
Tencent released Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters, licensed under Apache 2.0. The model outperforms similar-size models and rivals flagship open-source models with 2-5x parameters, supporting a 256K context length. It is available for free on OpenRouter until July 21st.
Apple's third developer betas of its version 27 operating systems introduce a built-in Uniform Type Identifier (UTI) for Markdown: `net.daringfireball.markdown`, which conforms to `public.utf8-plain-text`. The author has updated their own recommendation to match this UTF-8 encoding requirement.
Backblaze has stopped backing up cloud storage services like Dropbox and iCloud Drive because they use Apple's File Provider APIs, making files appear local without being fully downloaded. Users of Maestral, a Dropbox client that syncs to a regular local folder, report Backblaze still backs up their Dropbox content. Dropbox's own old-school client mode may also work, but results are inconsistent.
John Gruber reflects on Allen Pike's argument that ChatGPT's Mac app excels with live document attachments — a native feature iOS can't replicate. He notes Anthropic's chief product officer Mike Krieger moved to head an internal lab team rather than focus on polished desktop tools.
Maestral, the open-source Mac Dropbox client known for its simple and lightweight syncing, has been retired. Developer Sam Schott archived the project in July 2026, citing lack of time and no longer using Dropbox himself. The current version will continue working for now but will not receive further updates or maintenance.
After 28 years, longtime columnist Jason Snell published his final column at Macworld. He reflects on starting at the publication in 1997, when Apple was near bankruptcy, Steve Jobs had just returned, and the company's survival was uncertain. Snell had been transferred to Macworld after his former magazine, MacUser, folded.
Amazon has been accused of using seller data to create competing Amazon Basics products. The article warns that using AI tools exposes company data to AI providers who could become competitors, citing Anthropic's Claude Design launching after its CPO left Figma's board.
C2PA digitally signs images with cryptographic authenticity metadata, but is currently ineffective because almost no human-taken photos are signed while most AI images are—making signature removal a trivial way to fake human origin. For the system to work, all cameras must sign by default and all platforms must preserve metadata, requiring years of effort and regulation.
The author redesigned his notes blog's "Shuffle" button to navigate to a dedicated /shuffle/ page, which shows a brief "Shuffling..." message before redirecting to a random note via JavaScript. This avoids injecting all note IDs into every HTML page, reducing build changes. He intentionally added a slight delay to mimic an old CD player's shuffle feedback.
A programmer shares daily notes from attending PLDI in Boulder, detailing workshop participation, collaborative coding on Datalog and Knuth-Bendix, meeting researchers, and social activities.
Day One, the journaling app for Mac and iOS, has introduced Daily Chat, a guided reflection feature that uses AI to help users talk through their day and shape their thoughts into a journal entry. The feature is designed to assist those who struggle with how to start writing.
Revisiting his 2018 piece, John Gruber says his concern about Electron harming native Mac apps was justified, but native Mac apps are now resurgent. He critiques Anthropic's Claude as an "Electron turd," noting the irony that Claude itself can generate quality native Mac code.
Fantastical 4.1.15 introduces Calendar Mirroring, a privacy-focused feature that lets users connect two separate calendars so events from one automatically appear on another, with the option to show full details or just mark time as "Busy." All event data stays on the device and is not sent to Flexibits servers.
The article chronicles the history of the Columbia River dams, the Bonneville Power Administration, and the Master Grid. It details the BPA's pioneering work in long-distance power transmission, carrier current communications, and protective relaying systems that shaped the modern US electrical grid.
The article argues that Claude for Mac is an Electron app because its lead developer, Felix Rieseberg, is a former Microsoft engineer known for porting Slack to Electron and has a professional history strongly tied to the Electron framework.
Cory Doctorow argues that understanding computers at multiple abstraction levels—from cardboard CARDiac to BASIC code to modern "vibe coding"—builds "legibility" and Fingerspitzengefühl essential for problem-solving, and that automated code tools serve as a funnel for cultivating deeper technical skills.
EveryMac.com, a site providing detailed information on every Mac model, celebrated its 30th anniversary in 2026. The milestone is noted alongside Daring Fireball's 24th anniversary, with DF's John Gruber reflecting on EveryMac's longevity, having launched in 1996 before Steve Jobs returned to Apple.
Meta's engineering culture is deteriorating under Mark Zuckerberg and Scale AI CEO Alexandr Wang, who have introduced keyboard tracking, reassignments to data labeling, and AI-centric performance metrics. Critics argue this incentivizes performative AI use, drives away experienced engineers, and contributed to a major Instagram hijacking incident caused by AI-written and AI-reviewed code.
MG Siegler was banned from WhatsApp for the third time in three years without explanation, shortly after claiming a username. He criticizes Meta's unreliable appeal process and notes that banned accounts still receive messages undetected by senders.
Hackers have been stealing Instagram accounts by asking Meta's AI support bot to link a target account to a new email address, without any verification needed. The AI sends a code to the attacker's email, which is then used to reset the password and gain full access. One victim only recovered his account after sending Meta's AI a link to a news article about the vulnerability.
Cory Doctorow argues that AI coding paradoxes resolve when distinguishing "today's task" code from "accretive work." Drawing on the concept of "canonization," he explains that while personal, disposable code has value, the AI industry's focus on reverse centaurs consumes software canon without replenishing it.
The EU AI Act requires AI text outputs to be watermarked by August 2026. Methods like Google's SynthID and Unicode homoglyphs exist, but all text watermarks can be trivially removed via paraphrasing or character replacement, making them fundamentally different from image watermarks.
Social
No quotes found.