The official @Alibaba_Qwen account announced the release of Qwen3.8-Max and stated that its weights will be open-sourced next week, alongside the release of the developer-friendly Qwen3.8-27B. @alibaba_cloud’s concurrent introduction positions it as a work partner covering programming, reasoning, research, writing, and AI Agents. The key point is not a single parameter scale, but the simultaneous advancement of “flagship capability + locally deployable small models”: the former determines the ceiling, while the latter affects the coverage of the developer ecosystem. The open-source timeline and model names mentioned are official public information; community judgments on pricing, capability, and real-world experience are based on blogger tests and cannot replace official metrics.<\/p>\n
Sources:<\/p>\n
@cellinlab integrated the model into Claude Code, starting from an empty repository to build a functional 3D town with buses, malls, cinemas, tasks, and UI. Their records show the model can handle cross-file dependencies, review browser interfaces, and continue modifications based on visual feedback, maintaining established rules well during long tasks. The actual run also exposed issues like vehicle overlap, camera occlusion, floor position misjudgment, and ticketing sequence problems. 3D collisions, aesthetics, and extreme states still require manual, round-by-round review. @LufzzLiz’s comparative tests also noted significant improvements in the official version over the Preview on tasks like SVG, particle gravity, and black holes. However, these are visible blogger tests and should not be written as the model having achieved a “human-free” conclusion.<\/p>\n
Sources:<\/p>\n
@MiniMax_AI publicly announced H3 is available; but visible tests distinguish between “open-source” and “universally deployable.” Records compiled by @LufzzLiz show that a quantized version has been run on a setup with 12GB VRAM and 32GB RAM for a 480p\/20-step case, but a ~5-second video still took nearly 9 minutes. The public validation configuration for the unquantized version is two 32GB GPUs plus about 384GB of RAM. @Gorden_Sun also reminded home users of the video generation threshold, citing the 65GB model size and at least an A100 GPU. Additionally, a statement by @RyanLeeMiniMax, relayed by @MaxForAI, indicated that due to generated video copyright lawsuits, the license imposes regional restrictions on the US, EU, UK, and South Korea. This is a license specification and should not be simplified to mean the model files cannot be distributed.<\/p>\n
Sources:<\/p>\n
@xiaohu summarized that ByteDance released the Seedance 2.5 Prompt Guide and User Guide, covering single-sentence generation, scheduling with up to 50 reference materials, segmenting 30-second long videos, local edits, as well as prompt formulas and case studies for different scenarios. For creators using Dreamina or the API, the value lies in organizing trial-and-error experience into replicable templates. The current visible evidence is the blogger’s summary of the official documentation; specific effects still need to be verified with actual materials and account scenarios.<\/p>\n
Sources:<\/p>\n
@Chengzilhy, with 10 years of experience in women’s fashion retail, believes that fashion e-commerce sells not just clothes, but also aesthetics, scenarios, and the imagination of “looking better,” so AI-generated models themselves may not be a hindrance. A Xiaohongshu (Little Red Book) account case he relayed has about 2,000+ followers, with most videos consistently exceeding 10,000 views, and the store also has sales. His own solution involves quickly producing videos using one image and a prompt, and he plans to recruit creators with a foundation to jointly test traffic. The view counts, sales, and cost conclusions here are based on blogger observations and summaries and are not equivalent to universal platform patterns.<\/p>\n
Sources:<\/p>\n
According to @xiaohu’s summary of an internal Anthropic demo, Claude in Slack can proactively read discussions, chime in without being mentioned, then query data across Datadog and BigQuery, summarize solutions, generate UI sketches, and finally pull designers into a multi-person conversation. Another summary relayed Boris’s statement: internally at Anthropic, an average of about 90% of code is now written by Claude Code, shifting team roles more towards prototyping, building, maintaining, extending, and wrapping up. The “90%” figure and demo details are the blogger’s summary of public videos or statements; a more cautious judgment is that AI’s unit of work is shifting from single-turn answers towards sustained goals, tool permissions, and team collaboration loops.<\/p>\n
Sources:<\/p>\n
@imwsl90 shared the process of gradually building a management center starting from a personal homepage, product, and course websites: the system includes revenue analysis, channel data scraping, website monitoring, tool sites, and content creation, with plans to integrate more workflows into the system. @FuSheng_0306 documented a smart irrigation case: using AI on-site to run through a recommendation system demo in about 15 minutes, while clearly stating it’s not yet ready for direct deployment. Both pieces point to a practical boundary: AI can significantly shorten the validation and prototyping phases, but truly reusable gains come from solidifying personal business processes, data, and acceptance criteria into proprietary systems; demo speed does not directly equate to production delivery speed.<\/p>\n
Sources:<\/p>\n
@joshesye proposed having stronger models handle planning, task breakdown, and quality checking, while delegating bulk execution to cheaper, faster models; he believes this division of labor also applies to AI video. @Khazix0918’s experience post listed Qwen3.8-Max’s price as $2 per million input tokens and $6 per million output tokens, and stated its daily task and development experience is close to Kimi K3, but these are personal experiences and price quotes. The reusable conclusion is not that one model permanently wins, but rather separating directional judgment, execution, and review, then configuring models based on task value.<\/p>\n
Sources:<\/p>\n
Stats: Timeline Scans=360 Matched Bloggers=37 Matched Tweet Total=201 Weighted Tweet Score=159.45 Original Tweet Count=90 RT Tweet Count=40 Crawl Attempts=2 Boundary Coverage Status=tail_confidently_crossed_target_boundary<\/p>\n","protected":false},"excerpt":{"rendered":"
Alibaba Cloud releases Qwen3.8-Max and will soon open-source its weights. MiniMax H3 is open-source but has regional and hardware restrictions. Claude demonstrates a shift from single-turn dialogue to continuous team collaboration.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[19],"class_list":["post-1543","post","type-post","status-publish","format-standard","hentry","category-brief","tag-x--ai-"],"_links":{"self":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1543","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1543"}],"version-history":[{"count":0,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/posts\/1543\/revisions"}],"wp:attachment":[{"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1543"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1543"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/en.blog.liu-qi.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1543"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}