---

I. 3 Key Updates

Key Update 1｜DeepSeek V4-Flash-0731 Official Version Launches Public Beta — Lightweight Version Surpasses Own Flagship

On July 31, DeepSeek announced via API update logs that the official version of V4-Flash (version number V4-Flash-0731) launched public beta. Architecture unchanged (284B total parameters / 13B active, 1M context), purely relying on post-training to explode Agent capabilities: Terminal-Bench 2.1 score 82.7, approximately 21 points higher than the preview version, comprehensively surpassing its own V4-Pro preview version (72.1). API price remains at input $0.14 and output $0.28 per million tokens, weights synchronized open-sourced under MIT license. Simultaneously, native support for Responses API and Codex adaptation for the first time—developers can set DeepSeek as the model provider directly in Codex CLI and VS Code's Codex plugin. V4-Pro official version expected to release in early August.

Large Model Expert｜ Without changing architecture or touching parameters, pure post-training gained +21 points on Terminal-Bench, indicating that much of the improvement space for Agent capabilities lies in training data and alignment phases, rather than stacking parameters. Running an intelligence index of 50 with a 284B/13B MoE at this price point further lowers the unit cost of "good enough intelligence." V4-Pro official version hasn't come out yet; commercializing small models first is a very pragmatic strategy.

Software Expert｜ If you are using DeepSeek API for coding or running automation tasks: No code changes needed, model name remains deepseek-v4-flash, backend automatically upgrades, recommended to use directly, getting a free upgrade. Note two things: ① App and web versions haven't updated, regular chat users won't notice; ② Old interfaces deepseek-chat / deepseek-reasoner will be discontinued on October 24, don't delay migration scheduling.

---

Key Update 2｜ByteDance Seedance 2.5 Released — Single Generation 30 Seconds, Jimeng Membership Limited-Time 40% Off

On July 31, ByteDance officially released video generation model Seedance 2.5: Single generation duration doubled from 15 seconds in 2.0 to 30 seconds, supporting multi-turn continuation to produce coherent videos lasting minutes; Reference material limit increased to single instance 30 images + 10 videos + 10 audios; Added white model, green screen reference, and local video editing (precise timestamp control); Natively supports over 10 languages. Already online on Jimeng AI, Doubao Professional Edition, Coze, Xiaoyunque, API coming soon to Volcano Ark. XCMG Group (industrial training videos), XPeng Motors (product design visualization), etc., have confirmed integration. Jimeng AI simultaneous limited-time offer: Continuous yearly/monthly subscriptions 40% off, quarterly 30% off (July 31 – August 6), basic membership 79 yuan/month. Note: Seedance 2.5 generation credit consumption has significantly increased compared to 2.0 (30 seconds 720P requires 780 credits).

Large Model Expert｜ The core of 2.5 isn't the number "30 seconds," but the consistency engineering of long narratives—characters, scenes, and cameras don't collapse during multi-turn extensions, plus timestamp-level local editing, video generation moves from "lottery-style generation" to "modifiable, iterable production tools." XCMG and XPeng integrated on the release day, indicating it targets not just content creation, but industrial scenarios.

Software Expert｜ Users making short videos or e-commerce promotional videos: Jimeng's 40% off window lasts until August 6, try it this week if interested. But calculate first: 2.5's credit unit price is approx. 85% higher than 2.0, 30 seconds costs 780 credits, basic membership gives 725 credits per month, barely enough for one long video—casual hobbyists save money with 2.0, heavy commercial users grab membership during discount.

---

Key Update 3｜MiniMax H3 Video Model Online — 2K + Native Stereo Sound, And Soon to Be Open-Sourced

Same day (July 31), MiniMax released video model H3: Native 2K (2560×1440) output, 4–15 second clips, dialogue, ambient sound, music, and visuals generated in a single inference pass (not post-dubbing); "All-capable reference" system consumes up to 9 images + 3 videos + 3 audios per instance; Supports instruction-based local editing and motion transfer; 2K pay-as-you-go $0.13/second (one 15-second clip approx. $1.95), officially claimed to be less than one-third of mainstream models; Independent evaluation Artificial Analysis video editing leaderboard #1. Weights promised to be open-sourced "within days"—will become the first top-tier video model with open weights. Shadow side: Hailuo AI platform still faces copyright lawsuits from Disney, Universal, Warner Bros.

Large Model Expert｜ H3's engineering highlight is joint generation of audio and video in the same latent space, lip-sync, sound effects, and visuals naturally synchronize, this path has a higher ceiling than "generate first then dub." If weights are released as promised, video generation will replicate the open-source shockwave in the LLM field—the premium logic of closed-source video models will be rewritten.

Software Expert｜ Teams making ads or e-commerce product videos: Recommended to immediately try small-cost orders, 2K at $0.13/second is a genuine lowest price on the web, and supports local editing of "change product, keep scene," exactly the rigid demand for e-commerce scenarios. But two notes: ① Weights aren't truly released yet, don't finalize self-hosting plans; ② Commercial teams should let legal review copyright lawsuit risks before signing annual frameworks.

---

II. 10 Curated Updates

• Claude Opus 5 Tops Dual Programming Leaderboards (Voting cutoff Aug 1) — Arena WebDev (1702.9 points) and Image-to-WebDev (1668.6 points) double champion. For paying users writing code, Claude remains the quality ceiling currently.
• OpenAI GPT-5.6 Two Models Price Cut (July 30) — Luna slashed 80% to $0.20/$1.20 per million tokens, Terra slashed 20%; subscription prices unchanged. API call costs drop again, ChatGPT subscription users unaffected.
• ByteDance Major Org Adjustment: Feishu Merged into Doubao (July 30) — Feishu product team merged with Doubao, former Feishu head Xie Xin reports to Doubao head Zhao Qi; Sales merged into Volcano Engine. Feishu users unaffected short-term, long-term Doubao AI capabilities will accelerate injection into office suite.
• Tesla China Car System Officially Integrates Doubao Large Model (July 31) — Tesla owners' voice interaction experience is about to see a perceivable upgrade.
• DeepSeek Old Interface Retirement Countdown — deepseek-chat / deepseek-reasoner will be discontinued on October 24. Developers schedule migration in Q3 plan.
• Claude Sonnet 5 Discount Pricing Ends August 31 — Input price will rise from $2 to $3/million tokens (+50%), new tokenizer will burn 10–35% more tokens. Batch tasks completed within this month are more cost-effective.
• MiniMax HK Stocks Surge Intraday (July 31) — Stock price rose approx. 13%–17% on H3 release day. Market voted yes for the "open-source video model" story.
• Suno Loses Lawsuit in Germany (July 31) — Court ruled unauthorized use of GEMA library for training, must disclose income and compensate. Users making commercial background music with Suno, watch copyright compliance winds.
• Kimi K3·Max App Agent Trial Quota Updated August 2 — Users logging into Kimi today can check their new quota.
• Thinking Machines Open-Sources Inkling Small (Late July) — 276B total params / 12B active, Apache 2.0 license, performance approaching own large model. Self-hosting players get another free option.

---

III. Quick News Bar

• DeepSeek V4-Flash-0731 weights uploaded to Hugging Face, Unsloth quantized version 4-bit requires approx. 168GB memory.
• Caixin: Nvidia, Meta, Microsoft, Dell, etc. jointly support open-source models; Anthropic refuses to sign.
• Amazon invested $13.7 billion in OpenAI Series C preferred stock in Q2.
• vLLM announces zero-day support for Kimi K3, including KDA prefix caching and DSpark speculative decoding.
• Artificial Analysis: DeepSeek V4-Flash-0731 single-task cost is still approx. 60% lower than discounted GPT-5.6 Luna.
• Meta discloses future expenditure commitments near $700 billion, involving AI data centers and cloud computing.

---

IV. Tomorrow's Focus

• DeepSeek V4-Pro Official Version — Officially stated "release ASAP," expected early August, accompanying Harness tools
• MiniMax H3 Weights Release — Promised "within days," open-source community holding breath
• Seedance 2.5 API — Coming soon to Volcano Ark, enterprise users can start applying
• Jimeng 40% Off Window — Ends August 6
• Claude Sonnet 5 Price Increase — Ends August 31, 29 days remaining

---