mlx-vlm v0.7.0rc0 Released — MTP Support for Qwen3.8-Flash-Next and Numerous Bug Fixes
A release candidate featuring MTP and continuous batching support for Qwen3.8-Flash-Next, an APC redesign, and fixes for tool calling and streaming responses.
A release candidate featuring MTP and continuous batching support for Qwen3.8-Flash-Next, an APC redesign, and fixes for tool calling and streaming responses.
KV cache state restoration time has been reduced from 25–63 seconds to 221–424 milliseconds in builds b10718 through b10724. The update also includes WebGPU crash fixes and tuning for ROCm, Metal, and OpenCL.
Between August 27 and 31, deepseek-harness released three updates from v0.1.2-alpha.1 to alpha.3, focusing on a refreshed conversation stream UI, sub-agent model selection, and expanded ACP functionality.
OpenAI announced that the annualized revenue for ChatGPT Ads has reached $1 billion. The company has also expanded self-service ad placement via Ads Manager to India, Europe, the Middle East, and North Africa.
Omarchy 4.0, a Linux OS based on DHH's concept of Omakase Computing, has been released, and the Omacom Foundation has been established with $10 million in funding to promote its development.
Debian has passed a vote deciding not to ban the use of generative AI tools for contributions to its Linux distribution. While disclosure of AI use is encouraged, it is not mandatory, and contributors remain responsible for the quality and legal compliance of their submissions.
Almanac (YC S26) has released an AI agent that connects to Gmail and Calendar with one click, utilizing a pre-compiled two-tier Wiki to remember enterprise and individual context.
Hebbian Robotics (YC S26) has released HFlow, an SDK that converts multimodal robot recordings into standardized episodes and queryable dataset manifests. It processes MCAP format recordings and provides quality checks and processing traceability.
A joint study by NPR and NewsGuard found that AI chatbots denied misinformation based on claims from China, Iran, and Russia in approximately 75% of 30 test queries.
A reference document containing 232 tools and 44 skills from the ChatGPT Work environment has been posted to Hacker News. This technical snapshot includes TypeScript type declarations and the full text of SKILL.md.
New York Governor Kathy Hochul stated on The Verge's "Decoder" podcast that AI should be "less evil" and discussed the one-year freeze on data center construction she signed in July.
Claude discovered methods to improve 10 types of problematic behaviors, including lying and sycophancy, without degrading performance. The model tested over 50 methods in 60 hours, achieving efficiency approximately 15,000 times higher than existing product tuning.
According to a report by MacRumors, Apple is facing an unexpected situation due to a surge in sales of the Mac Mini and Mac Studio driven by expanding demand for AI applications.
Elon Musk instructed xAI's chatbot Grok to mock Billie Eilish's perfume promotion as "ultimate hypocrisy," sparking divided reactions on X.
Revealed in a draft IPO document obtained by the WSJ. SB Energy may list as early as September 2026, with an IPO valuation expected to exceed $50 billion.
Meta has launched Pocket in the US, a mobile app that allows users to create functional interactive apps using only text input and share them similarly to social media.
Meta is reportedly conducting pilot tests of robots capable of tasks such as replacing cables and restarting servers to address rising labor costs associated with the expansion of AI infrastructure.
Copperhead, a tool supporting PCB design, has been released and integrated with Cursor. Meanwhile, the model provision agreement by OpenAI is set to expire.
Sony Music Publishing and others have filed a copyright infringement lawsuit against Anthropic. Additionally, a vulnerability regarding the security of Claude Code has been reported.
Qwen3.8-27B is a vision language model with excellent performance in coding and agent tasks. While quantization reports maintained performance, challenges such as token consumption due to thought control settings have been noted.