---

I. 3 Key Updates

Key Point 1｜MiniMax H3 Officially Open-Sourced Today at Midnight — First Top-Tier Video Model with Open Weights

At midnight Beijing Time on August 3, MiniMax H3 was officially open-sourced on the ModelScope community, becoming the first top-tier video generation model with open weights, and also MiniMax's first open-source multimodal model (previously, three generations of M-series text models were open-sourced). Simultaneous progress: H3 official API pricing was announced on August 1—$0.081/second for 2K resolution (approx. 0.8 RMB/second, measured by Beijing News), with the 768p version coming soon at $0.047/second, half the price of mainstream 720p offerings. Channels are fully deployed: OpenRouter (Day 0 support), Venice, HailuoAI, MiniMax Official API, and domestic RunningHub launched it first. The model design initially considered compatibility with multiple domestic chips.

Correction Note: Yesterday, this column cited third-party sources stating H3 pricing was $0.13/second. Based on the official API pricing page from August 1, it should be $0.081/second. Correction hereby issued.

LLM Expert｜ Two technical points worth remembering for the industry: ① H3 systematically transferred methods validated in the text model era (complex problem decomposition, Reasoning, Scaling Context, Muon optimizer) into multimodal training. This "text methodology transfer" route may have become the new paradigm for video models; ② Native compatibility with multiple domestic chips + open weights equals handing domestic computing platforms a ready-made knife—private deployment stories for platforms like Ascend now have their best carrier starting today.

Software Expert｜ Three types of people should act today: ① Teams with GPU resources—download weights immediately and try self-hosting; 2K video generation now has zero marginal cost; ② Individual creators without graphics cards—go to RunningHub or Hailuo AI and pay-as-you-go directly. $0.081/second means a 15-second 2K video costs about 12 RMB, two orders of magnitude cheaper than outsourcing; ③ Those doing e-commerce/advertising—"change product, keep scene" partial editing + this price suggests small-cost trial orders this week. One caution: Copyright lawsuits like Disney's are unresolved; have legal review commercial annual contracts before signing.

---

Key Point 2｜First Wave of August Model Retirement Tide — Claude Opus 4.1 Shuts Down Day After Tomorrow (Aug 5)

Anthropic's Claude Opus 4.1 (claude-opus-4-1-20250805) will officially retire from the API on August 5, with migration target Opus 4.8 (pricing significantly reduced from $15/$75 to $5/$25 per million tokens). This is just the first domino in the August retirement wave: On August 10, OpenAI's gpt-5.2-chat-latest, gpt-5.3-chat-latest, and Google embedding-2-preview shut down on the same day (migrating to gpt-5.5 / gemini-embedding-2); On August 17, the entire Google imagen-4.0 series retires; On August 20, DeepSeek V4-Flash/V4-Pro preview versions on Azure platform retire; On September 24, the entire OpenAI Sora 2 series shuts down. Also note: Starting from Opus 4.7, passing non-default values for temperature/top_p/top_k will directly result in a 400 error.

LLM Expert｜ The retirement tide is essentially a byproduct of compressing the model iteration cycle from "years" to "months": Vendors no longer have the patience to maintain model matrices beyond three generations. The profound impact on the industry is—the era of hardcoding model IDs in code is over; "model routing layer + aliases + fallback" is moving from best practice to survival necessity. Anthropic's approach of announcing 60 days in advance is an industry benchmark, but not every company provides this buffer.

Software Expert｜ If your company has code calling Claude/OpenAI/Gemini APIs, do one thing today: Have engineers globally search code for strings like claude-opus-4-1, gpt-5.2, gpt-5.3. Anything not replaced by August 5 will start throwing errors and halting operations that day. Regular app users are unaffected, no need to panic. During migration, conveniently remove parameters like temperature from the code (control via prompts instead), otherwise switching to the new model will still cause errors.

---

Key Point 3｜Seedance 2.5 Enterprise Channel Opens — Volcano Ark Model Page Public, API Countdown Begins

Three days after ByteDance Seedance 2.5 release, Volcano Ark has published the model details page, officially confirming "online experience and API calls coming soon." The list of enterprises confirmed to integrate on release day continues to grow: XCMG Group (industrial training SOP videos), XPeng Motors (product design visualization), Agibot, Qiongche Intelligence (embodied intelligence training data synthesis), Lingchu Intelligence, Weifen Zhifei, Xspark AI, etc. The industrial background disclosed by Caixin is worth reading in contrast: Seedance 2.0 overseas call ratio has risen from 1/3 to 1/2, holding approx. 80% share in the AI short drama market; ByteDance's large model business ARR reached $4 billion, exceeding the total of other domestic model companies. On the individual side, the Jimeng 40% off membership window closes on August 6, 3 days left.

LLM Expert｜ Note the composition of the integration list: Robotics companies (Agibot, Qiongche, Weifen Zhifei) are more active than film/video companies—they don't want to "make videos," but use video generation for embodied intelligence training data synthesis. Video models are transforming from content tools to physical world simulators; this is industrial proof of the "world model" route. An ARR of $4 billion exceeding the domestic peer total indicates ByteDance's competitors are no longer domestic.

Software Expert｜ Enterprise users: You can submit API trial applications to Volcano Ark right now. First-batch channels usually come with discounts and support; waiting a month means a different price tag. Individual users, don't get distracted; you have only one decision window: Jimeng 40% off ends August 6. Heavy video creators must decide within these three days; light users can continue using the free quota for 2.0, no need to chase the new.

---

II. 10 Curated Updates

1. MiniMax H3 Launched on Venice (Aug 1): A privacy-focused video generation channel, officially stating "open weights will put more control in the hands of the community." → Adds a compliance option for teams sensitive to data privacy.

2. RunningHub First Launches H3 (Jul 31): Domestic AI creation platform integrated immediately. → Domestic creators who don't want to fuss with APIs can use it via web browser.

3. Kimi K3 Weights Open Source First Week Fermentation (Supplement · Jul 27): Hugging Face downloads and developer tests are hot, performing well on programming and frontend leaderboards; large-scale commercial use requires authorization. → For self-hosted text models, K3 remains one of the top domestic choices currently.

4. ByteDance Large Model Business ARR $4 Billion (Supplement · Jul 30): Exceeds total of other domestic model companies; Doubao daily token call volume 180 trillion. → Domestic large models entering a "one super, many strong" landscape, procurement bargaining room is shrinking.

5. OpenAI Sora 2 Entire Series Shuts Down Sept 24: Official retirement table confirms, no migration version. → Workflows still using Sora 2 must switch solutions within two months (Seedance/H3 are natural alternatives).

6. Google imagen-4.0 Entire Series Retires Aug 17: Migrating to gemini-3.1-flash-image. → Image generation pipelines using Imagen 4.0 must switch within two weeks.

7. Azure Platform DeepSeek V4-Flash/Pro Preview Versions Retire Aug 20. → Enterprises using DeepSeek via Azure, switch to formal endpoints.

8. n8n 2.27.0 Pre-release: Execution data can be stored in S3-compatible storage, OpenTelemetry management UI launched, GitHub/Kafka/OneDrive nodes enhanced. → Heavy users of self-hosted automation workflows worth upgrading; reserve several minutes for migration for large instances.

9. H3 is MiniMax's First Open Source Multimodal Model (Official statement Jul 31): Previously open-sourced three generations of M-series text models. → MiniMax's open source strategy expands from text to multimodal, clear ecosystem signal.

10. Product Hunt Aug 2 Rankings: AI workflow automation and AI dictation tools continue to dominate charts, new product PraiseEngine (AI customer review collection) launched. → Overseas small tool innovation is still finding opportunities in "workflow segmentation," while domestic homogeneous products are crowded.

---

III. Flash News Bar

• Seedance 2.0 overseas call ratio rose from 1/3 to 1/2 (disclosed by ByteDance on July 30).
• Seedance holds approx. 80% market share in AI short dramas (Caixin).
• ByteDance average monthly burn rate basis shows large model business ARR at $4 billion (Caixin).
• Weekly LLM Digest: China's open source camp is becoming the "price and weight rule maker" (80aj Weekly #38).
• Volcano Ark Coding Plan 25% off ongoing, compatible with coding tools like Claude Code (xmsumi).
• Beijing News measurement: H3 generation pricing 0.8 RMB/second (2K), consistent with official $0.081/second basis.

---

IV. Tomorrow's Focus

1. August 5 (Day After Tomorrow): Claude Opus 4.1 retirement—last migration window for API fixed ID users;

2. August 6: Jimeng 40% off membership window closes, 3 days left;

3. August 10: OpenAI gpt-5.2/5.3-chat-latest, Google embedding-2-preview retire;

4. August 17: Google i