AI Model Daily - 2026-08-04 | QName AI

30-Second Summary

Today's main event is Alibaba Qwen's official release of Qwen3.8-Max: the official blog calls it the most capable Qwen model to date and it is now available through Qwen Cloud / Alibaba Cloud APIs; the official X post explicitly says open weights for Qwen3.8-Max and Qwen3.8-27B will arrive next week, so it should not be described as "weights already released" today.
Arena and Artificial Analysis now provide independent signals: Arena WebDev lists qwen3.8-max at #4 with 1668±18; Artificial Analysis lists a release date of 2026-08-03, Intelligence Index 53, 1M context, and $2 / $6 per 1M input/output tokens.
The Qwen3.8-Max capability narrative is centered on coding and cowork/office agents: official and media materials repeatedly mention 2.4T total parameters, 95B active parameters, 1M context, long-horizon tasks, and multimodal input; AA also marks the current model as proprietary, meaning the weights are not public yet.
MiniMax H3 and HappyHorse1.1 are near-window carryover signals: MiniMax's official release notes date H3 to 2026-07-31; Alibaba Token Plan places HappyHorse1.1 in the same access bundle as Qwen3.8-Max, DeepSeek V4, and GLM-5.2, but HappyHorse1.1 launched before this window.
The Google/Nano Banana signal is not a new model launch; it is the reported rollback/pause of Nano Banana image generation inside Google Earth. Gemini Omni is also not a new launch today and should be tracked as product availability and risk-governance context.

Must Watch

1. Qwen3.8-Max: Qwen's Flagship Officially Ships, API Live, Weights Still Next Week

Event type: Official release / X first signal / access update / open-weight promise to verify.
First/official/availability time: Qwen's official blog and @Alibaba_Qwen released Qwen3.8-Max on 2026-08-03; the official X post says open weights will be released next week and Qwen3.8-27B will also become open weights.
Capability change: Qwen says Qwen3.8-Max is its strongest model to date, aimed at coding and cowork. Public materials and media coverage focus on 2.4T total parameters, 95B active parameters, 1M context, text/image/video input, and long-horizon coding/office workflows. Artificial Analysis records it as a reasoning model with text/image/video input, text output, and 1.0M context.
Access: Qwen Cloud / Alibaba Cloud Model Studio / Alibaba Cloud Token Plan; Arena links to qwencloud.com; Artificial Analysis currently tracks only Alibaba Cloud as a provider.
Impact: This is the real new-model story today. Qwen has moved from preview/open-weight anticipation to a purchasable API plus third-party leaderboard entries. But official weights, license, and HF/ModelScope/GitHub repositories are not live yet, so self-hosting evaluation has to wait for next week.
Sources: Qwen official blog, Alibaba_Qwen X, QwenCloud model link from Arena, Alibaba Cloud Token Plan, Artificial Analysis Qwen3.8 Max.

2. Arena: Qwen3.8-Max Enters the WebDev Front Pack

Event type: Arena appearance / leaderboard change / X first signal.
First/official/availability time: Arena's official X account posted on 2026-08-03 that Qwen3.8-Max landed on the Frontend Code Arena; the public WebDev leaderboard page is dated 2026-08-01 but already includes qwen3.8-max.
Capability change: On WebDev Overall, qwen3.8-max ranks #4 with 1668+18/-18, 1,563 votes, $2 / $6 pricing, and 1M context. It sits behind claude-opus-5-max, kimi-k3-max, and claude-opus-5-high. In the lab view, Alibaba ranks #3.
Access: Arena WebDev leaderboard and QwenCloud.
Impact: This is more useful than vendor self-ranking. Qwen3.8-Max is not #1 overall, but it has already moved above GPT-5.6 Sol xHigh on the WebDev track, making it a serious candidate for coding agent and frontend generation tests.
Sources: Arena X, Arena WebDev leaderboard, Arena WebDev labs view.

3. Artificial Analysis: Independent Qwen3.8-Max Metrics Are Live

Event type: Public leaderboard / analysis-source update.
First/official/availability time: Artificial Analysis lists Qwen3.8 Max with release date 2026-08-03.
Capability change: AA reports Intelligence Index 53, output speed 46.5 tokens/s, TTFT 2.48s, input price $2.00/M, output price $6.00/M, 7:2:1 cache/input/output blended price $1.18/M, and context window 1.0M tokens. AA also notes the model produced 150M output tokens during Intelligence Index evaluation, far above the comparable median of 63M.
Access: Alibaba Cloud provider.
Impact: AA's score is more conservative than some social-media "second-best/top-tier" narratives, but the price and 1M context make it highly relevant for high-cache, high-context agent loops. Key tests: whether verbosity increases real cost, whether tool use is stable, and whether extended thinking hurts end-to-end latency.
Sources: Artificial Analysis model page, Artificial Analysis provider page.

4. MiniMax H3: Not a Today Launch, but Still on the Open Video-Model Track

Event type: Official release / media coverage / near-window carryover.
First/official/availability time: MiniMax API release notes list MiniMax H3 on 2026-07-31; Chinese AI news/media continued to surface H3 multimodal generation and open-model notes on 2026-08-03.
Capability change: MiniMax describes H3 as a new-generation open general-purpose multimodal video model that understands creative intent across text, image, video, and audio context and produces more natural, coherent video generation and expression.
Access: MiniMax API docs and MiniMax Hub.
Impact: Do not mark today as H3's launch day. Still, H3 belongs on the same video-generation watchlist as HappyHorse, Gemini Omni, Kling, and Seedance; next checks should cover Hugging Face/GitHub weights, commercial API, resolution, audio-video sync, reference-image/video control, and pricing.
Sources: MiniMax model release notes, MiniMax Hub, Sohu / Yicai AI brief.

5. Nano Banana / Gemini Omni: Not a New Model, but a Product-Risk Signal

Event type: Product access update / media coverage / risk-governance signal.
First/official/availability time: Google launched Google Earth + Nano Banana image generation on 2026-07-30; multiple outlets reported between July 31 and August 3 that the feature was pulled or paused because of misuse and fake geospatial imagery risks. Gemini Omni was released before this window.
Capability change: Google's launch post said Nano Banana in Google Earth could generate custom images grounded in satellite, aerial, and 3D imagery. Gemini Omni's official page continues to emphasize "reference anything" and text/image/video/audio input into a unified output path, starting with video generation and editing.
Access: Gemini image generation, Gemini Omni / Google Flow, and Google Earth status subject to the actual official Google surface.
Impact: This is a textbook case of strong generative media entering a trusted geospatial context and creating risk. Product implication is more important than raw model capability: map, news, finance, legal-evidence, and other evidentiary image workflows need stronger provenance, watermarking, share limits, and abuse monitoring.
Sources: Google Earth Nano Banana launch, Mobile World Live rollback report, Gemini Omni model page, Introducing Gemini Omni.

Arena Watch

Arena Text / Software & IT Services: the public page lists qwen3.8-max at #9 with 1520±16 Preliminary, 1,405 votes, $2 / $6, and 1M context. It is above nearby kimi-k3-max (#12) and near gpt-5.6-sol-xhigh (#11), but still behind Claude Fable/Opus and Meta Muse Spark 1.1. Source: Arena text leaderboard.
Arena WebDev: qwen3.8-max ranks #4 with 1668+18/-18 Preliminary and 1,563 votes; the top three are claude-opus-5-max, kimi-k3-max, and claude-opus-5-high. This track matters most because Qwen's release narrative emphasizes coding/cowork. Source: Arena WebDev leaderboard.
Artificial Analysis: Qwen3.8 Max has Intelligence Index 53, release date 2026-08-03, 1M context, $2/$6 input/output pricing, and output speed 46.5 t/s. AA currently says weights are not publicly available, so the official "open weights next week" line remains a future promise. Source: AA model page.
MiniMax H3: no new independent H3 Arena entry was seen today; the official release notes date is 2026-07-31. Source: MiniMax release notes.
Conclusion: today's leaderboard story is Qwen3.8-Max. Its WebDev result is strong enough to enter first-tier agentic coding tests; whether it can replace Kimi K3 / Claude / GPT-5.6 depends on real-task verbosity, tool-call stability, and end-to-end time.

X Watch

Chrome logged-in state pending: this run connected to the Chrome extension as requested, but X search pages repeatedly hit Chrome-control load/control timeouts, so stable bulk reading of visible article elements was not possible. This report therefore uses publicly crawlable X result snippets and official posts first, without reading cookies, local storage, passwords, or session files.
@Alibaba_Qwen: the core original post is the Qwen3.8-Max release. It explicitly says open weights are coming next week and mentions Qwen3.8-27B as open-weight too; it also carries pricing narrative such as $2 input, $6 output, and $0.25/M cache. Trust: high for official access and announcement; the actual weights still need repository verification next week. Source: Alibaba_Qwen X.
@arena: posted that Qwen3.8-Max reached #4 on Frontend Code Arena with score 1,668. Trust: high as official leaderboard signal; still verify against the live leaderboard because scores keep moving. Source: Arena X.
@xiaohu / XiaoHu daily: public search summaries show XiaoHu highlighted Alibaba's Qwen3.8-Max, calling it the strongest Qwen model, 2.4T parameters, and able to autonomously run for 16 days. The "already open-sourced" wording conflicts with Qwen's official "open weights next week" timing, so treat it as a Chinese auxiliary source requiring verification. Sources: xiaohu X profile/search, XiaoHu.AI Daily.
@alibaba_cloud: Alibaba Cloud Token Plan continues to bundle Qwen3.8-Max, HappyHorse1.1, DeepSeek V4, and GLM-5.2; HappyHorse1.1 is not a today launch. Sources: Alibaba Cloud Token Plan, Alibaba Cloud HappyHorse X.
@OpenAI, @OpenAIDevs, @AnthropicAI, @GoogleDeepMind, @GoogleAI, @GeminiApp, @xai, @grok, @AIatMeta, @MistralAI, @Kimi_Moonshot, @deepseek_ai, @runwayml, @LumaLabsAI, @Kling_ai, @cohere: public official sources did not show a high-confidence new foundation-model launch in the window; visible changes were mainly existing models, product access, policy/safety discussion, or near-window carryover.

Pending Confirmation

Qwen3.8-Max open weights: official wording says "next week"; confirm exact date, repository, license, file scale, quantized variants, whether Qwen3.8-27B is included, commercial use terms, and whether the Max-class open weights match the API model.
Qwen3.8-Max parameter accounting: media and official-channel summaries repeatedly mention 2.4T total / 95B active, while Artificial Analysis says Alibaba has not disclosed model size or parameter count. Future reports should prefer explicit fields from the official blog/model card and avoid treating media repetition as independent evidence.
Qwen3.8-Max pricing: official X and AA both show $2/$6, but cache hit, explicit cache create/read, Token Plan credit conversion, international vs China-region prices, and third-party gateway markups need separate cost math.
Qwen3.8-Max real agent capability: Arena WebDev #4 is a strong signal, but local tests should cover multi-file edits, long repository reads, tool-call JSON stability, excessive output, and low-human-intervention continuity.
MiniMax H3 open-weight status: official notes say "open general-purpose", but confirm weight/code download location, license, inference framework, commercial limits, and API pricing.
Nano Banana Google Earth rollback: many media reports exist; continue watching whether Google publishes an official incident/postmortem, relaunch timing, SynthID/visible watermark/share-limit changes, and abuse-control policies.
Gemini Omni / HappyHorse1.1: no new launch in this window. Future tracking should record official first appearance, global availability, API availability, and first leaderboard appearance separately to avoid mistaking secondary spread for a launch.