All AI model reports
Collected in Hong Kong time

AI Model Daily - 2026-05-23

2026-05-23 AI model news: Qwen3.7-Max is today's top model signal, moving from soon to available through OpenRouter and Qwen Cloud API access.

Collected at: 2026-05-23 06:31 HKT
Collection method: Chrome session, web verification, public leaderboards, and official sources

30-Second Summary

  • The most actionable model signal today is Qwen3.7-Max: it moved from "coming soon" to "API entry available." OpenRouter lists qwen/qwen3.7-max, while Qwen Cloud shows API examples, pricing, and a 1M context window.
  • The highest-signal update from @xiaohu remains NetEase Youdao Confucius4 / Confucius4-TTS: the multimodal model weights are visible, while the TTS repository is open but code and weights are still in preparation.
  • Gemini Omni continues to circulate across X and Xiaohu AI Daily. Google positions it as "any input -> video," but API access is still described as coming in the next few weeks, and it has not appeared on the Artificial Analysis video leaderboard yet.
  • Command A+ has a fresh distribution signal: after Cohere's official launch, Microsoft Azure reshared that Command A+ is available in Microsoft Foundry as a Managed Compute offer.
  • Video arenas did not materially reshuffle: Dreamina Seedance 2.0 720p and HappyHorse-1.0 remain the leading T2V/I2V references.

Must Watch Today

1. Qwen3.7-Max

  • Event type: official launch, high-signal X activity, and API availability.
  • Official timing: Alibaba Cloud Summit announcement on 2026-05-20; Qwen's official X account pinned Qwen3.7-Max on 2026-05-21.
  • Availability: OpenRouter lists qwen/qwen3.7-max, released May 21, 2026; Qwen Cloud lists qwen3.7-max with OpenAI-compatible API examples.
  • Capability shift: positioned for agent-centric workloads, especially coding, office/productivity, and long-horizon autonomous execution. Qwen Cloud marks text input/output, function calling, cache, structured outputs, batches, and web search support.
  • Pricing/context: OpenRouter lists $2.50 / $7.50 per 1M, with 1M context. Qwen Cloud lists 1M context, about 991.80K max input and 65.53K max output.
  • Leaderboard status: qwen3.7-max-preview ranks 14th on LMArena Text Overall, Elo 1475 ±10, 3,741 votes, Preliminary.
  • Impact: this is today's most important "test now" model signal. Prioritize internal coding/agent benchmarks against Gemini 3.5 Flash, GPT-5.5, and Claude Opus 4.7.
  • Sources: Qwen X, OpenRouter, Qwen Cloud, Alibaba Cloud, LMArena

2. NetEase Youdao Confucius4 / Confucius4-TTS

  • Event type: X first-signal plus open model card/repository confirmation.
  • First signal: @xiaohu posted roughly 16 hours ago that NetEase Youdao open sourced two Confucius4 models, one for mathematical visual reasoning and one for voice cloning.
  • Availability: the multimodal model is visible on Hugging Face; the TTS repository is published on GitHub.
  • Capability shift: Confucius4 is based on Qwen3.5-27B, Apache 2.0, and positioned for multimodal reasoning, with emphasis on visual math benchmarks and shorter CoT outputs. Confucius4-TTS targets 14 languages, cross-lingual zero-shot TTS, and speaker identity preservation.
  • Key limitation: the Confucius4-TTS README says "Code and model weights are currently under preparation," so it should be treated as demo-first for now, not fully local-runnable open source.
  • Impact: this is today's strongest open-model signal. Start with small visual math reasoning samples for the multimodal model, then evaluate TTS after weights land.
  • Sources: Xiaohu X, Confucius4 Hugging Face, Confucius4-TTS GitHub

3. Gemini Omni / Gemini Omni Flash

  • Event type: continued activity after official launch.
  • Official release date: 2026-05-19.
  • Capability shift: Google says Gemini Omni Flash can create content from any input, with the first version starting from video. It accepts image, audio, video, and text inputs, then generates video grounded in Gemini world knowledge.
  • Availability: Gemini app, Flow, Google AI Plus/Pro/Ultra, and product-side access. Official API availability is still described as coming in the next few weeks.
  • X/aggregator signal: Xiaohu AI Daily led with Gemini Omni on 2026-05-21; live X search still shows many demos and discussions, but the noise level is high.
  • Leaderboard status: no Gemini Omni entry was found on the Artificial Analysis video leaderboard in this pass. Video comparisons should still use Veo 3.1, Kling 3.0 Omni, Seedance, HappyHorse, and similar references.
  • Impact: watch API timing and independent video arena entry. For now, do not judge real model strength from demos alone.
  • Sources: Google Blog, Xiaohu AI Daily

4. Cohere Command A+

  • Event type: official launch, open weights, and distribution expansion.
  • Official release date: 2026-05-20.
  • Capability shift: Cohere calls it the fastest and strongest Command model, using an MoE architecture with 218B total / 25B active, 128K input context, and support for text, images, tool use, reasoning, and 48 languages.
  • Availability: Cohere says weights are downloadable from Hugging Face under Apache 2.0. On X, Microsoft Azure reshared that Command A+ is available in Microsoft Foundry as a Managed Compute offer.
  • Leaderboard status: Cohere cites an Artificial Analysis Intelligence Index score of 37, positioning it above other leading open models.
  • Impact: a candidate for private deployment and enterprise agentic workloads, especially as a complement to Qwen open models, Mistral 3, and Gemma 4.
  • Sources: Cohere Blog, Cohere X

5. Gemini 3.5 Flash

  • Event type: official launch and LMArena leaderboard confirmation.
  • Official release date: 2026-05-19.
  • Capability shift: Google positions it for complex agentic workflows, with emphasis on coding, long-horizon tasks, and action-taking.
  • Leaderboard status: gemini-3.5-flash ranks 9th on LMArena Text Overall, Elo 1480 ±8, 5,907 votes, Preliminary.
  • Availability: Gemini app, AI Mode, Gemini API, AI Studio, Antigravity, and enterprise platforms.
  • Impact: currently Google's most practical coding/agent comparison model, second only to today's newly available Qwen3.7-Max signal.
  • Sources: Google Blog, LMArena

Arena Watch

LMArena Text Overall

  • Leaderboard date: 2026-05-18.
  • Scale: 6,297,180 votes, 360 models.
  • Key changes:
    • gemini-3.5-flash: 9th, Elo 1480 ±8, 5,907 votes, Preliminary.
    • qwen3.7-max-preview: 14th, Elo 1475 ±10, 3,741 votes, Preliminary.
    • gemma-4-31b: 40th, Elo 1451 ±8.
  • Read: Qwen3.7-Max and Gemini 3.5 Flash are both near the top tier, but sample sizes remain relatively small. Watch how rankings stabilize after Preliminary status clears.
  • Source: LMArena Text

Artificial Analysis Video Arena

  • Text-to-video with audio:
    • Dreamina Seedance 2.0 720p: 1st, Elo 1213, 8,209 samples.
    • HappyHorse-1.0: 2nd, Elo 1212, 6,869 samples, API Coming soon.
    • Veo 3.1: 3rd, Elo 1100.
  • Image-to-video with audio:
    • Dreamina Seedance 2.0 720p: 1st, Elo 1183.
    • HappyHorse-1.0: 2nd, Elo 1166.
    • Veo 3.1 Fast: 3rd, Elo 1094.
  • Watch point: HappyHorse-1.0 remains the key semi-anonymous video model signal. It is shown as Alibaba-ATH, released Apr 2026, with API still unavailable. Gemini Omni has not appeared on this leaderboard yet.
  • Sources: Artificial Analysis T2V, Artificial Analysis I2V

X Watch

  • @xiaohu: the high-signal model item is Confucius4 / Confucius4-TTS. Recent posts from the past two hours mainly discuss X Premium authorization for Openclaw/OpenCode/Hermes and Grok usage as product distribution, so they are not included in today's core model list.
  • @Alibaba_Qwen: pinned Qwen3.7-Max; a repost from 12 hours ago confirms OpenRouter availability, which is the key signal moving Qwen from launch to usable API.
  • @cohere: pinned Command A+; Microsoft Azure reshared availability in Microsoft Foundry, which is a distribution-channel signal.
  • @ArtificialAnlys and @GoogleDeepMind: Chrome pages only loaded shells in this pass; official pages and Artificial Analysis leaderboards were used instead.
  • X live search: the f=live search page only returned filters and shell UI, with no usable post body. Marked as Chrome session follow-up needed. Account pages were more reliable than search pages.

To Confirm

  • Gemini Omni API: Google still says access is coming in the next few weeks. Check Gemini API, AI Studio, Vertex AI, and Flow again tomorrow.
  • Gemini Omni leaderboard: no independent Artificial Analysis video ranking was found yet. Continue watching T2V/I2V arena inclusion.
  • HappyHorse-1.0: leaderboard shows Alibaba-ATH, released Apr 2026, API Coming soon, but there is still no formal official model card or API documentation.
  • Confucius4-TTS: GitHub repository exists, but code and weights are still in preparation. A live demo is not the same as fully runnable open source.
  • Qwen3.7-Max: OpenRouter and Qwen Cloud list availability, while Alibaba Cloud's official news page still says worldwide soon. Run practical API checks for stability and region/account restrictions.

Suggested Next Steps

  • Test Qwen3.7-Max first today: coding agents, multi-file edits, long-context tool use, cost, and response stability.
  • Run 5-10 visual math samples on Confucius4 to see whether it really saves tokens versus same-scale Qwen3.5-27B.
  • Benchmark Command A+ on enterprise documents, RAG, and tool calling, as a private-deployment complement to Qwen3.7-Max closed API use.
  • Keep Gemini Omni in watch mode for now; wait for API access or independent leaderboard entry before systematic evaluation.