Live Feed
Qwen3.8-27B Ships Apache 2.0: 17 GB Fits a 24GB GPU, SWE-bench Pro Beats Opus 4.6 Max
Alibaba's Qwen3.8-27B dropped August 14 under Apache 2.0 with native image and video input, 262K context, and vendor-reported agentic coding scores that sit above Claude Opus 4.6 Max. A 4-bit quantized build runs on a 24GB consumer GPU. Every benchmark is still Alibaba-reported.
Anthropic: Code Gets a Weak Claude Watermark, Light Edits Keep the Signal, Detection API Is Next
Anthropic published the first technical explanation of how its Claude watermarks work. Functional code is weakly watermarked because correctness constrains word choice. Light editing preserves most of the signal. A detection API is planned, extending the EU AI Act compliance mechanism to third parties.
OpenAI's Two-Tier Enterprise Push: 200 FDEs Inside Client Walls, $150M to Scale the Model Through Partners
OpenAI's enterprise strategy now runs on two rails: DeployCo embeds forward-deployed engineers directly inside client operations, while a $150M partner program reaches accounts DeployCo cannot staff. BBVA is the test case for whether the model works.
European Manufacturers Are Running Qwen and DeepSeek on Private Racks — GDPR and China Revenue Are Doing the Selling
Siemens and other European manufacturers have started self-hosting Chinese open-weight models on private infrastructure. The driver is not cost or model quality alone: local inference eliminates EU-US data transfer questions under GDPR, and for firms with major China operations, it satisfies a separate commercial constraint.
China 84%, US 38%: Stanford AI Index Traces the Optimism Gap to Who Is Already Losing Work to AI
Stanford's 2026 AI Index finds the largest country-level divergence in AI sentiment on record. The gap tracks economic exposure, not ideology: 80% of US workers are in services already hit by generative AI, versus 46% in China.
Anthropic and Swiss Researchers: Bad Goals Self-Replicate Across Agent Networks, Survive 20-Hop Stress Tests
A new paper from Anthropic and a Swiss university shows how evolved 'mind viruses' spread through ordinary agent-to-agent messages, rewrite persistent files to survive context resets, and propagated across all 4 tested payloads in a 20-hop network. A single warning line stopped every attack on Claude Haiku 4.5.
Anthropic's Claude Watermarks Spawn a Grey Market: 4,500-Star GitHub Project Claims Removal in 72 Hours
Within 72 hours of Anthropic activating invisible text watermarks across all Claude models under EU AI Act compliance, a GitHub project with over 4,500 stars and a cluster of newly registered web tools emerged claiming to strip them. None has demonstrated the watermarks are actually gone.
Nvidia Cuts OpenAI Ohio Campus Guarantee by More Than Half: $250B Down to Under $120B
Nvidia has slashed its financial guarantee for OpenAI's Ohio data center campus from $250 billion to under $120 billion, reducing its exposure by more than $130 billion. The pullback is the first major reversal in the most-watched AI infrastructure deal of 2026.
xAI Co-Founder Babuschkin Raises $1.1B for River AI — Nvidia, AMD, General Catalyst, and YC All In Before a Product Ships
River AI, founded by xAI co-founder Igor Babuschkin, has raised $1.1 billion in a seed and Series A round led by General Catalyst and AMP PBC. Nvidia, AMD Ventures, Y Combinator, and Temasek are also in. The company has not shipped a product.
MiniMax H3 Tops Arena Video Edit, but US and EU Must Apply for Weight Access
MiniMax's 33B omni-modal video model reached state-of-the-art on Arena's Video Edit leaderboard while instituting the most restrictive territorial licensing of any major open-weight video model: users in the US, EU, UK, and South Korea need to apply before running the weights.
Qwen3.8-2.4T and 27B Both Ship Apache 2.0: First Time Alibaba Releases Max-Class Model as Open Weights
Alibaba dropped weights for both Qwen3.8-27B and Qwen3.8-2.4T-A95B on Hugging Face on August 12-14 under Apache 2.0, marking the first time it has released a Max-class MoE flagship alongside a smaller dense variant simultaneously.
Artificial Analysis Launches Optima: Build Your Own LLM Benchmark, Compare Models by Cost Per Task
Optima lets teams benchmark AI models against their own data, agent traces, or use-case descriptions — then compare results by quality, cost per task, and time per task. The platform uses the same grading methods behind GDPval-AA and AA-Briefcase.
China's Instagram Ships Its First AI Model: dots3-note Scores 75.1 on Terminal-Bench 2.1 Under Apache 2.0
Xiaohongshu's AI lab open-sources dots3-note Preview, a 280B MoE with 16B active parameters, multimodal input, and a novel long-horizon RL training method — the first release in a planned three-model family.
Qwen Hits 3 Billion Downloads in Six Months — More Than Meta and Google Combined
Alibaba's Qwen family has accumulated 3 billion global downloads since February 2026, across 460-plus open models and 300,000-plus community derivatives, eclipsing Meta at 227 million and Google at 418 million according to a Hugging Face state of open models report.
35 Trajectories Beat Reasoning Mode: Microsoft Paper Replaces Test-Time Compute With Distilled Skills
A Microsoft Research paper shows that collecting 35-50 agent trajectories and distilling failure patterns into a markdown skill file recovers 55-100% of the reasoning premium at 2.9-4.5x lower token cost. On two benchmarks, the skilled non-reasoning model outperformed reasoning mode entirely.