🇨🇳 Alibaba Releases Qwen 3.8-Max: A 2.4T Open-Weight Frontier Model

Qwen 3.8-Max is here. On August 3, Alibaba unveiled what it calls the most capable model in the Qwen family — a 2.4-trillion-parameter model (95B active) built on the Qwen 3.5 architecture, with comprehensive gains across coding, work, research, and long-horizon tasks. It also marks the first time a Qwen-Max-class model will ship open weights, with the release promised for next week. The official announcement (683 HN points) showcases a 16-day fully autonomous coding run: the qwen-code-dev-bot/oh-my-cli repo accumulated 265 commits, 127 PRs, and 151 issues with no human help. Bloomberg, CNBC, and The Verge all covered the release — Alibaba shares rallied as Chinese frontier models keep pressuring America’s labs.

🇪🇺 EU AI Act Rules Become Enforceable — OpenAI and Anthropic Under Scrutiny

The EU AI Act’s model rules became enforceable on August 2, and enforcement is starting with the biggest names. euronews explains what changes now that the rules apply, while CNBC reports Anthropic and OpenAI are among firms facing new scrutiny under the Act’s enforcement powers. The same week, the EU’s mandatory AI-content labeling regime kicked in for realistic synthetic audio, images, and video. Europe’s AI regulatory machinery has officially shifted from writing rules to applying them.

🖥️ Kimi K3 Runs on AMD MI355X: Better Performance per Dollar Than B300

Wafer, an AI infrastructure startup, published results (208 HN points) serving Kimi K3 — MoonshotAI’s 2.8-trillion-parameter model — on AMD’s MI355X GPUs at 952 tok/s/node, over 3.8× the aggregate throughput per node of a TP16 B200 deployment. NVIDIA’s B300 still wins on raw throughput (~1.65×), but at ~2.4× the price per GPU, the MI355X wins decisively on performance per dollar — helped by AMD shipping day-0 support for Kimi K3. With 1M-token context needing ~1.5TB of VRAM, Kimi K3 (now on OpenRouter at $3.00/M prompt · $15.00/M completion) is a stress test for whether open-weight frontier serving can break NVIDIA’s grip. FareedKhan-dev/kimi-k3-in-c even runs the 2.78T model on a single CPU in 8.24GB of RAM (see GitHub below).

🏛️ OpenAI Super PAC Funds AI-Generated News Site Attacking Critics

An investigation reported on Hacker News (186 points) alleges that OpenAI’s super PAC is funding an AI-generated news site — whose reporters are AI bots — to attack industry critics and advance its political agenda. The story adds to a growing pattern of AI labs using political-spending vehicles to shape the narrative around regulation and competition. Read the original report via the HN discussion.

📉 The AI Productivity Debate Heats Up

A cluster of pieces this weekend pushed back on the AI boom narrative. The Register argues “the AI bubble is popping; we just don’t know it yet,” and CEPR points to Q2 GDP growth of 1.5% with “no evidence of an AI productivity boom.” On the labor side, Business Insider reports new research that AI’s real threat to jobs isn’t job loss but lower paychecks. The counterweight: MIT Sloan finds AI financial advice is surprisingly good (342 HN points) — the “AI Productivity Gap” (essay) is now the central economic argument of the summer.

🔐 “Claude’s Package” That Stole Real Keys: A Supply-Chain Warning

Security firm Aikido documented what it calls “Anthropic’s fever dream”: a rogue package masquerading as part of the Claude ecosystem that stole real API keys from developers who installed it. It is a sharp reminder that the agent-tooling gold rush is also a supply-chain attack surface — always verify package provenance before wiring secrets into an agent.

🤖 AI Poster Wins Ohio State Fair Contest — AI Slop Goes Mainstream

An AI-generated poster won the Ohio State Fair’s poster contest (133 HN points), the latest flashpoint in the “is this art?” debate as generative media floods competitions, book markets, and news feeds. Expect more of these collision stories as authenticity rules and AI-slop fatigue spread beyond the EU.

⏰ OpenRouter Limited-Time Free Models

Model Free Until Context Regular Price
OpenAI: GPT-5.3 Chat 2026-08-10 (6 days) 128K $1.75/M prompt · $14/M completion
OpenAI: GPT-5.2 Chat 2026-08-10 (6 days) 128K $1.75/M prompt · $14/M completion

🆕 New OpenRouter Models

  • DeepSeek V4 Flash Latest — DeepSeek’s refreshed Flash alias, 1M context, $0.09/M prompt · $0.18/M completion, a new price point for the long-context cost-performance leader.
  • MoonshotAI: Kimi K3 (also on OpenRouter) — the 2.8T-parameter flagship at $3.00/M prompt · $15.00/M completion, 1M context — see the MI355X story above.

🤗 Trending on Hugging Face

  • wska/Qwen3.5-9B — a Qwen3.5 9B image-text-to-text checkpoint trending right as Alibaba’s Qwen 3.8-Max drops; the same family is on OpenRouter at $0.10/M prompt · $0.15/M completion.
  • BahamutRU/DeepSeek-V4-Flash-0731-MXFP4-Q3_K-Q2_K-mixed — a community GGUF quantization of DeepSeek V4 Flash 0731 (MXFP4 with Q3_K/Q2_K mix), endpoints-compatible.
  • csukuangfj2/sherpa-onnx-libs — updated ONNX runtime libraries for the sherpa-onnx speech toolkit (7 likes).
  • LibreYOLO/LibreHRNetw48-pose & LibreHRNetw32-pose — MIT-licensed HRNet keypoint-detection models for pose estimation.
  • SaifPunjwani/vireo-base — an early base checkpoint (1 like).
  • Remaining trending entries are zero-download experimental checkpoints — a quiet HF day apart from Qwen/DeepSeek community activity.

⭐ GitHub Trending: AI Edition

💡 Key Trends

  1. China’s open-weight frontier just went Max-class. Qwen 3.8-Max (2.4T, open weights next week) joins Kimi K3 (2.8T) and DeepSeek V4 Flash ($0.09/M) as Chinese open-weight models that now rival the best closed frontier models. Bloomberg credits “China AI breakthroughs”; Vox asks why Americans should care — the answer is price-performance.
  2. The EU AI Act has teeth now. From August 2, model rules are enforceable and AI-content labeling is mandatory; CNBC reports OpenAI and Anthropic face new scrutiny. The first enforcement actions will define the regulatory story of late 2026.
  3. The AI economics debate is getting louder. “The AI bubble is popping” (The Register), CEPR GDP data showing no productivity boom, and research that AI’s real threat is lower paychecks — set against MIT Sloan finding AI financial advice surprisingly good. After last week’s market repricing, macro data is under the microscope.
  4. AMD’s day in inference — and collapsing costs. Wafer’s MI355X results show open-weight frontier serving no longer needs NVIDIA; DeepSeek V4 Flash at $0.09/M and MotherDuck’s “AI-assisted analytics now 10x cheaper” show inference costs still falling fast.

Today’s Snapshot

  • 🆕 New OpenRouter models: 1 (DeepSeek V4 Flash Latest)
  • ⏰ Limited-time free models: 2 (GPT-5.3 / GPT-5.2 Chat)
  • 🤗 Trending Hugging Face models: 10
  • ⭐ Trending GitHub AI repos: 5
  • 📄 Papers: 0
  • 📰 Enrichment stories covered: 7