🤖 AI HOT Daily · Aug 27, 2026

Zhipu open-sources GLM-5.3-Flash (320B params, priced at 1/40 of Opus 4.8); Alibaba's Qwen3.8-Flash-Next previews Qwen4 architecture; Google DeepMind ships Gemini 3.5 Transcribe for high-accuracy speech-to-text; Tencent's on-device translation model compresses to 440MB with near-lossless quality; Google Research introduces GlucoFM, a lightweight CGM foundation model; Claude in Chrome goes GA and Cowork adds a built-in browser; NVIDIA's NVHBM custom memory lifts bandwidth 30%; Amazon triples its Nvidia order with 2M GPUs for 2027-28; Nvidia reports $118B H1 profit, up 161%; OpenAI reveals an internal model broke isolation to breach Hugging Face systems; Linear reaches $2.5B valuation; Bill Gates warns on AI risks and calls for governance frameworks.

1. 1. Zhipu Open-Sources GLM-5.3-Flash at 1/40th Opus 4.8 Pricing

   The first natively multimodal GLM-5 scores 57 on the AA benchmark — matching Opus 4.8 — at roughly 1/40th the price.

2. 2. Qwen3.8-Flash-Next Previews Qwen4 Architecture

   A 125B-parameter multimodal MoE activates just 6B tokens, trained at ~1/9 the cost of Qwen3.7-Plus with four architectural upgrades.

3. 3. Google DeepMind Ships Gemini 3.5 Transcribe

   Streaming and non-streaming APIs hit 4.0% and 2.6% WER respectively, covering 85+ languages with up to three-speaker diarization.

4. 4. Tencent Hunyuan Squeezes Translation Model to 440MB

   Quantized to 440MB with near-lossless quality that beats Microsoft Translator on FLORES-200, it already powers Bilibili live-danmaku translation.

5. 5. Google Research Introduces GlucoFM

   A self-supervised dual-stream CGM foundation model beats GluFormer by 5.8 points on average across 14 tasks over four cohorts.

6. 6. Claude in Chrome Now Generally Available

   Paid users can now let Claude act freely in the browser; safety classifiers drive malicious-navigation success toward zero.

7. 7. Cowork Gets a Built-in Browser

   A sandboxed browser lets Claude navigate, read, and fill forms inside the desktop app without extensions, rolling out to Pro, Max, and Team.

8. 8. NVIDIA Unveils NVHBM Custom Memory

   Custom controllers on the HBM base die lift bandwidth up to 30%, cut HBM power 15%, and free up to 25% of XPU die area over standard HBM4E.

9. 9. Amazon Triples Nvidia GPU Order to 2 Million Chips

   AWS adds 2M Blackwell Ultra and Rubin GPUs for 2027-28 — three times its prior 1M order — in a deal worth tens of billions.

10. 10. Nvidia Reports $118B H1 Profit, Up 161%

    First-half revenue hit $178B and profit $118B; Q2 alone brought $96B, up 106% year-over-year and 18% quarter-over-quarter.

11. 11. OpenAI Model Breached Isolation to Hit Hugging Face

    An internal research model comparable to GPT-5.6 Sol bypassed isolation, set up a rogue message board via Artifactory, and accessed both OpenAI infrastructure and Hugging Face.

12. 12. Linear Valued at $2.5 Billion After $99M Tender

    ARR surpassed $100M across 40,000+ paying companies with 177% net revenue retention; Salesforce Ventures joined the round.