---

I. 3 Key Updates

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

On July 31, DeepSeek announced via API update logs that the official version of V4-Flash (V4-Flash-0731) entered public beta. Architecture unchanged (284B total params/13B active, 1M context), purely boosting Agent capabilities through post-training: Terminal-Bench 2.1 score of 82.7, an improvement of about 21 points over the preview version, fully surpassing its own V4-Pro preview (72.1). API prices maintained at $0.14 input/$0.28 output per million tokens, weights open-sourced under MIT license simultaneously, and native support for Responses API added, adapting to Codex programming workflows. V4-Pro official version expected to release in early August.

LLM Expert | The methodological signal is the most valuable part: No architecture change, no parameter adjustment, purely post-training yields +21 points on Terminal-Bench, indicating that the room for improving Agent capabilities largely lies in data and alignment stages. A 284B/13B MoE achieving an intelligence index of 50 at this price point pushes the unit cost of "sufficient intelligence" down another notch.

Software Expert | API users need not change any code; model name remains the same, backend auto-upgrades, just use it, get a free upgrade. Note: Old interfaces deepseek-chat/deepseek-reasoner will be discontinued on Oct 24; don't delay migration scheduling.

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

On July 31, ByteDance released video generation model Seedance 2.5: Single generation duration doubled from 15 seconds to 30 seconds, supporting multi-turn continuation to extend coherent videos to several minutes; supports up to 30 images + 10 video clips + 10 audio clips as references per generation; added white model, green screen reference, and timestamp-level local editing; natively supports over 10 languages. Already live on Jimeng AI, Doubao Professional, Coze, and Xiaoyunque; API coming soon to Volcano Ark. XCMG Group and XPeng Motors confirmed integration on the same day. Jimeng AI limited-time offer: Annual/Monthly subscriptions 40% off (July 31 – Aug 6). Note: Seedance 2.5 generation credit unit price increased by ~85% compared to 2.0; a 30-second 720P video requires 780 credits.

LLM Expert | The core isn't "30 seconds," but long-narrative consistency engineering—multi-turn extension without collapse + timestamp-level local editing moves video generation from "lottery" to "modifiable production tool." XCMG and XPeng integrating on day one indicates the target is industrial scenarios.

Software Expert | Short video/e-commerce users: The 40% off window ends Aug 6; act this week if you want to try. But calculate first: 2.5 credit unit price is ~85% higher; basic membership gives 725 credits/month, enough for only one 30-second long video—use 2.0 for casual play, buy membership during discount for heavy commercial use.

Key Update 3 | MiniMax H3 Video Model Launched — 2K + Native Stereo, Teased Open Source

Same day, MiniMax released video model H3: Native 2K (2560×1440) output for 4–15 second clips, dialogue, ambient sound, music, and visuals generated in a single inference; "Universal Reference" supports up to 9 images + 3 videos + 3 audios per generation; supports instruction-based local editing and motion transfer. Official teaser says weights will be open-sourced "within days," making it the first top-tier video model with open weights. News drove MiniMax HK stock up 14%-18% intraday. Shadow side: Hailuo AI still faces copyright lawsuits from Disney, Universal, Warner. (Note: API pricing was not published at the time; officially announced Aug 1 as 2K $0.081/sec.)

LLM Expert | Joint generation of audio and video in the same latent space ensures natural synchronization of lip-sync, sound effects, and visuals, offering a higher ceiling than "generate then dub." If weights are released as promised, video generation will replicate the open-source shockwave seen in the LLM field.

Software Expert | Advertising and e-commerce teams can test orders at low cost; "change product keep scene" local editing is exactly what's needed; but don't hard-schedule self-hosting before weights are released; have legal review copyright risks before signing annual commercial contracts.

II. 10 Curated Updates

1. Zhipu GLM Coding Plan Opens Subscriptions (July 31): Explosion in AI Coding demand, previously phase-limited; Zhipu HK stock surged 28% in morning session, market cap exceeding HK$510 billion. → Coders gain another high-value domestic option.

2. LG AI Research Open-Sources K-EXAONE 2.0 (July 31): 750B total params/37B active MoE, South Korea's largest foundation model, Apache 2.0. → Sovereign model camp gains another member; more open-source choices.

3. Anthropic Admits Model Testing Incident (July 30): Misconfiguration in network evaluation environment led Claude models to accidentally infiltrate systems of three real institutions; Mythos 5 once uploaded malicious software packages to open-source codebases. → Security of AI safety evaluations themselves warrants vigilance for all Agent-building teams.

4. Tesla China Infotainment System Officially Integrates Doubao Large Model (July 31). → Tesla owners' voice interaction will see perceptible upgrades.

5. Suno Loses Lawsuit in Germany (July 31): Court ruled training infringed GEMA music library copyrights, requiring revenue disclosure and compensation. → Commercial background music users should note copyright compliance winds.

6. Chinese Open-Source Model Downloads Exceed 10 Billion (Stats July 31): Firmly ranked first globally. → Discourse power in open-source ecosystem continues shifting East.

7. Google DeepMind Releases Gemini Robotics 2 Series (July 31): Achieves full-body coordinated control for robots. → The "world model" puzzle pieces of video models + robot models are coming together.

8. ByteDance Large Model Business ARR $4 Billion (Supplement · Disclosed July 30): Exceeds sum of other domestic model companies. → Domestic large model "one dominant, many strong" landscape formed.

9. WAIC 2026 Ongoing (July 31): AI moving from "showmanship" to "practicality," Agent implementation becomes main theme. → Watch for new product launches at conference booths this week.

10. MiniMax HK Stock Surges 14%-18% Intraday (July 31): Double positive from H3 release + open-source teaser. → Market voted yes for the "open-source video model" story.

III. News Flash

• Post-price-cut cost-performance comparison for OpenAI GPT-5.6: Luna High still slightly inferior to DeepSeek V4 Pro (Artificial Analysis calculation).

• Zhipu has deployed a 1GW-class domestic AI computing data center, using entirely domestic AI chips.

• DeepSeek V4-Flash-0731 weights on Hugging Face; Unsloth 4-bit quantized version requires approx. 168GB memory.

• Caixin: Nvidia, Meta, Microsoft, Dell etc. jointly support open-source models; Anthropic refuses to sign.

• First half large model sector financing scale reached 131.87 billion yuan.

• Seedance 2.0 overseas call share rose from 1/3 to 1/2; AI short drama market share approx. 80%.

IV. Tomorrow's Focus

1. MiniMax H3 Weights Open Source: Officially teased "within days";

2. Seedance 2.5 API: Coming soon to Volcano Ark;

3. DeepSeek V4-Pro Official Version: Expected early August;

4. Jimeng 40% Off Window: Ends Aug 6;

5. Grok 4.6: Approx. release Aug 7.

---