US federal agencies issued a joint advisory on Sept. 8 alleging that Chinese AI companies including Alibaba, DeepSeek, MiniMax, Moonshot AI, StepFun and Z.AI have run large-scale operations to pull outputs from US models like Claude, GPT, Gemini and Grok since late 2024. The advisory says these firms extracted billions of tokens through millions of queries, using techniques such as chain-of-thought extraction and automated evasion to dodge blocking measures, in order to train their own competing systems.
A follow-up experiment tested four open models—DeepSeek V4 Flash, Inkling, Kimi K3, and Qwen3.8 A95B—by inserting the first 1% of GPT-5.5 Pro's reasoning trace into each model's own reasoning channel before letting it generate answers freely. Researchers then measured how much of GPT-5.5 Pro's visible answer text overlapped with each model's output. Qwen3.8 showed the largest jump, with overlap rising from 33.92% unprefilled to 54.50% with the GPT-5.5 Pro prefill, a 20.58 percentage-point increase, while other models showed much smaller shifts.
CISA, the NSA, and the FBI issued a joint advisory naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI as having harvested billions of tokens from Anthropic, OpenAI, Google, and xAI models since late 2024. The firms allegedly used fraudulent accounts, proxy networks, and automated failover systems to bypass rate limits and extract restricted reasoning data at industrial scale.
DeepSeek confirmed it will officially release its V4.1 Flash model on September 10, 2026 (Beijing Time), stating internal testing shows it outperforms V4 Pro on performance, cost, speed, and task completion time. Until V4.1 Pro arrives, all Pro-model requests will be automatically redirected to V4.1 Flash and charged at the cheaper Flash pricing tier.
The NSA, CISA and FBI issued a joint advisory alleging that Chinese AI companies including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI extracted billions of tokens from US frontier models such as Claude, GPT, Gemini and Grok since 2024. The agencies say this data was used to train competing systems like DeepSeek's R1 and Moonshot's Kimi K2/K3, describing the practice as 'distillation activities at an industrial scale.'
The developer behind the iMessage AI assistant Olly, which has processed over 18 million messages with roughly a third routed through OpenRouter's open-source models, published a technical breakdown of problems encountered at scale. The core issue: OpenRouter can send identical model requests to any of about 20 different hosting providers, each running the same weights but with different precision, optimizations and tool parsers, producing measurably different benchmark results. For DeepSeek V4 Flash, GPQA Diamond and TAU-Bench Airline scores varied by several points across providers on the same day.
The NSA, CISA, and FBI issued a joint advisory alleging that Chinese AI companies, including DeepSeek and Moonshot AI, have run large-scale campaigns since 2024 to extract capabilities from leading US models like ChatGPT, Claude, Gemini, and Grok. The agencies say these firms used millions of queries routed through multiple accounts, APIs, proxies, and cloud providers to generate synthetic training data and build competing models such as DeepSeek's R1/R3 and Moonshot's Kimi-K2/K3.
A research team reports developing a pretraining method that matches DeepSeek V4 Pro Base's performance using roughly 50 times less compute, at an estimated cost of about $0.5 million on GB200 chips. Scaling the same recipe up 10x further, to roughly $4 million, reportedly surpassed all publicly available open base models on perplexity benchmarks.
A newcomer called StartLux, founded by veteran programmer Chen Danyan and formerly known as Yuandian Xinghui, placed second in China's CAICT MCP specialized benchmark with its StartLux-V1.0-27B-Preview model. The 27-billion-parameter model trailed DeepSeek-V4-Pro, a 1.6-trillion-parameter system, by just 1.3 percentage points, despite running locally on consumer-grade PCs rather than in the cloud. Chen has argued that local models will eventually outcompete cloud-based ones on efficiency and market share.
Writer Jordan Andersen expands on a post by Disesdi Shoshana Cox questioning why, despite AI supposedly boosting developer productivity tenfold, no wave of new billion-dollar startups comparable to Airbnb, Stripe, or Dropbox has emerged. Instead, the biggest winners of the generative AI boom are AI companies themselves—OpenAI, Anthropic, and High-Flyer—while broader effects include cluttered inboxes, spammy social content, and degraded search results.
Jerry James benchmarked an 8-GPU NVIDIA RTX PRO 6000 Blackwell system paired with an AMD EPYC 9555 CPU, offering 768GB of combined GDDR7 memory. Rather than splitting huge models like Llama-3.1 405B or DeepSeek-R1 671B across all eight PCIe cards, which introduces heavy latency, the team found the setup performs best running independent single-GPU model instances in parallel.
The vLLM project released version 0.28.0, combining 584 commits from 270 contributors. The update centers on deep performance work for Kimi-K3, including new decode context parallelism, fused kernels, memory-saving expert sharding, and ROCm support, alongside DeepSeek V4 improvements such as end-to-end sparse MLA and AMD Quark NVFP4 support.
GoKawiil