Skip to content

← X AI Daily

X AI Daily

Date17 November 2026 · TuesdayWindow2026-11-16 02:00 to 2026-11-17 07:58CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products and Applications

Alibaba releases Qwen4-Max general availability: 2.4 trillion total parameters, 71 billion active, official claim of roughly 60% lower average cost per task

At 10:01 Beijing time on November 16, the official account of Alibaba Cloud's Tongyi team, @Alibaba_Qwen, announced the general availability of Qwen4-Max. This is the GA release of the flagship model whose preview began on November 11: a new MoE architecture, native vision, and a 1 million token context. The official press release states 2.4 trillion total parameters, with 71 billion activated per token, roughly 25% fewer activated parameters than Qwen3.8-Max, which is also at the 2.4 trillion scale.

This is worth noting because the fourth-generation Tongyi flagship has moved from preview into general availability, and the official pitch is centered on using fewer tokens and lowering per-task cost, not just on another set of benchmark scores. In terms of reach, it was also the most-engaged item in this window: the official launch post had about 2.121 million views, 14.0 thousand likes, 1,507 reposts, 759 replies, 3,217 bookmarks, and 1,188 quotes; the follow-up post in the same thread discussing cost had about 114 thousand views and 698 likes. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the Alibaba Cloud press release confirms the November 16 release, 2.4 trillion total parameters, 71 billion active, the 1 million token context, and that on its in-house Work360 evaluation (covering 721 professional domains), the High reasoning tier used about 35% fewer tokens than Qwen3.8-Max and cut average cost per task by about 60%. The official thread added absolute figures from the same evaluation set: the average Work360 cost per task fell from $4.43 to $1.67; token usage dropped about 24% at the Xhigh tier and about 35% at the High tier, and about 48% and 49% respectively on Terminal-Bench 4.0. Availability covers QwenCloud's Token Plan and API, plus QwenWork and QwenStudio; the press release also says smaller open-weight models in the Qwen4 family will be released later. The official pricing page lists Singapore international deployment pricing at $1.50 per million input tokens and $4.50 per million output tokens, with implicit cache hits and explicit cache reads both at $0.075; thinking is billed as output. The QwenCloud model page and the pricing table allow these prices to be checked. The documentation also states that the qwen4-max-preview preview period runs from 10:00 on November 11, 2026 to 24:00 on November 25 (UTC+8), that the preview will be retired after it ends, and that the GA release is now the recommended version.

The official account also relayed several external leaderboard results: 6th on Code Arena WebDev, 74 points above Qwen3.8-Max; a 7.5% net win rate on Agent Arena, with Alibaba ranked 5th in the labs overall standings; and 9 points above Qwen3.8-Max on Artificial Analysis. These are the official account's attributions of leaderboard positions, not independent re-tests by this report.

Assessment: the GA specifications, availability, and price list can be checked. The cost and performance figures come mainly from the official in-house Work360 evaluation and from third-party leaderboards relayed by the official account. Discussion in Chinese and English within the window was not one-sided; some hands-on tests found almost no visible difference between the GA and preview versions, while others argued it still clearly trails frontier closed models.

Tencent releases Hy4 flash: 224 billion total parameters, 12 billion active, official claim that everyday-work blind tests came close to Claude Haiku 5.5

At 14:38 Beijing time on November 16, Tencent Hunyuan's official account @TencentHunyuan released Hy4 flash, positioned as a low-cost model for everyday productivity. The official press release states: 224 billion total parameters (including 64 billion Engram parameters), 12 billion activated per token, multimodal input support, and a context of up to 1 million tokens. The official launch post had about 536 thousand views, 2,475 likes, 219 reposts, 187 replies, 527 bookmarks, and 177 quotes. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the Tencent press release and the technical page confirm the above specifications and API pricing: $0.070 per million input tokens, $0.281 per million output tokens, and $0.003 per million cached input tokens. The press release also says: KV cache is about 1/24 that of Hy3 and 1/6 that of Hy4; reinforcement learning reduced average output tokens for end-to-end tasks by 50%; decoding speed exceeds 150 tokens per second; and on the internal WorkIndex benchmark, cost per task is about 1/13 that of Hy4 Preview, 1/20 that of Hy4, and 1/8 that of DeepSeek-V4.1-Flash-1017 (Max tier, peak pricing), with overall performance described as slightly better than the larger Hy4 Preview (770 billion total parameters, 49 billion active). In a blind test with 85 internal WorkBuddy users across 27 categories and 259 tasks, user satisfaction was close to Claude Haiku 5.5, and among the compared models it had the lowest cost per task and the shortest completion time. The official post framed the result as "on par with Claude Haiku 5.5 (high), with cost per task down to 1/18", while the press release used "close to". The model is already live on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis; at launch it said OpenRouter access was coming soon, and later the same day the official account confirmed it was live there with a limited-time 50% discount. From now through November 30, WorkBuddy is free to use and TokenHub is at 50% of standard pricing. The official recommendation is to use flash for everyday tasks and to keep using Hy4 for complex long-horizon work such as software engineering.

Assessment: this is a clear product launch, and the pricing, availability, and limited-time offers can be checked. The comparison with Haiku 5.5, and the cost multiples relative to DeepSeek Flash and Hy4, come from Tencent's internal evaluations and internal user blind tests, not from public third-party replication.

OpenAI announces first phase of embedded assessments, contracting with Apollo Research, METR, Transluce, Goodfire

At 00:03 Beijing time on November 17, OpenAI's official account announced that it has signed agreements with Apollo Research, METR, Transluce, and Goodfire to conduct the first phase of "embedded assessments", giving outside experts extended deep access to internal models and systems to assess alignment and whether safeguards are effective. The official post had about 604 thousand views, 2,729 likes, 265 reposts, 564 replies, 444 bookmarks, and 276 quotes. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the OpenAI official blog post confirms the division of work among the four parties: Apollo Research will assess whether models covertly act against developer intent and whether existing safeguards can detect it, and lists four claims to be tested; Goodfire is embedding interpretability researchers who will use its platform for training-process monitoring, debugging, and alignment, primarily serving internal researchers while publishing findings over time; METR plans two efforts, the first assessing misaligned behavior in what it describes as the largest reinforcement learning training to date, with the scope of the second monitoring evaluation still to be agreed by both sides; Transluce will first assess the quality and coverage of post-training alignment tests, and intends to run a broader follow-up evaluation of about eight weeks in early 2027. The blog post states that assessors can access checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and relevant researchers as needed. On the same day, Goodfire also published a collaboration note.

In the same window, The Information reported that OpenAI had held multiple talks this summer about acquiring Goodfire, an interpretability startup backed by Anthropic among others, with valuation discussions above $5 billion, and that no deal was reached. The discussion circulated somewhat on X, for example wallstengine's relay with about 91 thousand views and 281 likes. This report was unable to open the original The Information article; the acquisition details above are only a relay of that outlet's reporting and should not be treated as a completed or still-active transaction.

Assessment: the four-party collaboration is an officially confirmed safety evaluation arrangement, not a model release. The acquisition rumor and today's embedded collaboration announcement can be considered together, but they are not the same thing, and there is no official confirmation that the acquisition came close to closing.

Hugging Face researchers release PyBot: about 640,000 lines of pure-Python rules, self-tested to match SmolLM-360M-Instruct

Between 22:19 and 22:24 Beijing time on November 16, Hugging Face researchers Lewis Tunstall and Leandro von Werra introduced PyBot: a rule-based chatbot written in pure Python by more than 1,000 agents, with no weights and no training, using only if-else statements, regex, and lookup tables. Tunstall's intro post had about 17 thousand views and 175 likes; von Werra's intro post had about 64 thousand views, 607 likes, 56 reposts, 45 replies, and 560 bookmarks. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the GitHub repository description confirms that more than 1,000 agents wrote and evaluated over 5,000 versions, raising the win rate on SmolTalkBench (judged by a large model against reference answers) from 0% to 26%, comparable to SmolLM-360M-Instruct; the code is about 640,000 lines, plus about 550,000 lines of tests written by the agents themselves; it is licensed Apache 2.0 and runs with only Python 3.8 or higher, with no additional packages. The repository also provides terminal usage and a Hugging Face Space demo. The description does not say which model the agents used; Tunstall says in his post that each agent was an instance of DeepSeek-V4.1-Flash, and that the work was completed within a week.

Assessment: this is a reproducible research demonstration, asking "how far can a rule-based chatbot go if the rule-writing work is handed to agents", not a product benchmarked against frontier models. The 26% win rate comes from a project-built benchmark and large-model judging, and cannot be extrapolated to general conversational ability.

2. AI Compute and Industry

Databricks releases AI Index: open-weight models' share of enterprise tokens rises from 7.3% to 30.9% in about 3.5 months

At 01:15 Beijing time on November 17, Databricks engineering lead Patrick Wendell (@pwendell) announced the release of the Databricks AI Index. The data comes from more than 10,000 enterprises that manage model calls through Unity Gateway, spanning digital-native companies, Fortune 500, and Global 2000 firms. His announcement post had about 75 thousand views, 482 likes, 132 reposts, 21 replies, and 269 bookmarks. Colleague Yuchen Jin's summary post had about 7,783 views and 152 likes. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the Databricks official blog confirms that open-weight models' share of enterprise token traffic rose from 7.3% in July to 30.9% in November; the share of enterprises using only a single model vendor fell from about 64% in July to 28% in the week of November 2. Within proprietary models, in the six weeks after Claude Opus 5.5 launched, its share rose from 7.2% to 37.9% of proprietary-model tokens, and Anthropic's overall share rose from 57.6% to 68.8%. In coding evaluations, DeepSeek V4.1 and GLM 5.3 tied for 3rd with a 93.1% task success rate, and three open-weight models made the top 10. The blog also gives examples: GPT-6 Luna and Claude Opus 5.5 both have an 89.1% success rate on its coding benchmark, but average cost per task is $0.05 and $1.62 respectively; Gemini 3.8 Flash has a low unit price but an average of 140 turns pushed its cost per task to $11.97. The index also has an online page: Databricks AI Index.

Assessment: this is an observation based on traffic through Databricks' own gateway, not a full-industry census. The sample is sizeable and explicitly includes calls to external vendors such as Anthropic and OpenAI, but enterprises that choose to use Unity Gateway are themselves a biased group, and these figures cannot be taken directly as global enterprise market share.

Noeri, founded by Stanford professors Yejin Choi and Carlos Guestrin, raises $110 million and launches publicly

Stanford computer science professor Yejin Choi and Carlos Guestrin launched Noeri as co-CEOs, positioning the company as letting enterprises build, own, and keep improving specialized AI using their own data and processes. GeekWire reported on November 16 that the round was $110 million, led by Menlo Ventures, with participation from Madrona, Kindred Ventures, Wing, NEA, as well as Cisco Investments, NVIDIA, and Snowflake; the valuation was not disclosed. The most widely circulated item on X was DocumentingAGI's relay, with about 25 thousand views and 250 likes. Data was collected around 08:05 on 2026-11-17 (Shanghai time).

Verified: the company website confirms the two co-CEOs, the $110 million raise, and the product description of "learning in production, changing tools and models, and upgrading only after human approval"; the GeekWire report checked the investor list and states that Bloomberg was first to report the round that day and that the valuation was not public. Guestrin is a co-author of XGBoost, founded Turi (later acquired by Apple), and led machine learning at Apple; Choi is a MacArthur Fellow who previously did research at Ai2 and NVIDIA and is now a Stanford professor.

Assessment: this is a founder launch and funding announcement, not an evaluation of a product already deployed at scale. The company says customers retain ownership of the workflows and models they develop (subject to third-party model licenses), and the public materials so far contain no independently checkable customer outcome data.

3. Chinese-Language AI Ecosystem

Qwen4-Max GA and Hy4 flash were both Chinese vendors' releases that day, and their specifications are covered above; here we only add information that appeared in Chinese-language discussion.

For Qwen4-Max, the first reaction in Chinese-language circles was not only about official leaderboards. Some noticed that the GA release arrived almost quietly; for example, leo114119 (about 5,272 views, 56 likes) wrote that it was already usable in Qoder. A batch of comparisons with the preview version followed: the same author's front-end comparison post (about 5,018 views, 17 likes) said there was "almost no visible major change" from preview; another post argued that on some front-end tasks the GA version was still worse than the preview. There were also more positive immediate reactions, for example wanxubo (about 8,397 views, 39 likes) described it as "formally entering the second tier of Chinese models", while also saying the benchmark scores were unremarkable and it did not reach the Fable he had expected. These are individual hands-on tests or impressions, not a unified evaluation.

The new information about Hy4 flash on the Chinese product side concerns access and free trials: the domestic and international versions of WorkBuddy added it simultaneously, with free use from November 16 to 23:59 on November 30, while Hy4's free period ended and Hy3 will be retired in the app. For example, leo114119's intro post had about 2,478 views and 17 likes; engagement was low, but the content matches Tencent's official launch description.