AI 日报 — 2026-08-14

Google has introduced HEIR, an open-source compiler within its Private Computing Toolkit, designed to enable cryptographically-secure private AI inference. Since its announcement in 2023, HEIR has fostered collaborations with hardware accelerator developers like Belfort and Cornami, and academic institutions including Georgia Tech and Tsinghua University. This has led to several peer-reviewed publications and numerous citations, demonstrating its role as a productive research platform. Google aims to make homomorphic encryption easy to develop, fast to run, and ubiquitous across the industry.

This research introduces a contract-grade verifier for LLM-generated GPU kernels, detailed in a 17-page document with 3 figures. The paper, archived at doi:10.5281/zenodo.21563213 and arXiv:2608.12700 [cs.LG], explores subjects including Machine Learning, Hardware Architecture, and Distributed, Parallel, and Cluster Computing. Authored by Rishi Shah, this version (v1) was submitted on August 13, 2026.

Apple has reportedly trained a large language model (LLM) specifically for the China market, with support from Alibaba. This development would position Apple as the first foreign company to offer a proprietary AI model in China. The information comes from three individuals familiar with the matter, as reported by Reuters.

Z.ai has launched GLM-5.3, which utilizes the same foundational model as GLM-5.2. This new version incorporates scaled post-training to enhance its coding and cyber skills. Z.ai also announced plans to release the weights for GLM-5.3 within two weeks. The company previously developed IndexShare for efficient long-context processing and SAO for reinforcement learning on long-horizon tasks with GLM-5.2.

The Qwen 3.8 27B model repository on Hugging Face provides FP8-quantized weights and configuration files for a post-trained model in the Transformers format. It specifies parameters for an "Instruct" mode, including temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, and repetition_penalty=1.0. The model's output structure includes appending messages with roles like "assistant" and content fields such as "answer_content" and "reasoning_content."

Gemini 3.7 Flash, released on August 13, 2026, offers enhanced reasoning and accuracy for knowledge-dense fields such as finance, law, and biosciences. It significantly outperforms 3.6 Flash on the GDP.pdf benchmark, achieving 34.0% compared to 22.0%. Additionally, 3.7 Flash surpasses 3.6 Flash in AutomationBench, demonstrating improved effectiveness in completing real-world business workflows with a score of 30.4% versus 17.0%.

DeepSeek Harness (dsh), an open-source agent harness from DeepSeek AI, is now available in developer preview. It features a plugin-based architecture powered by Cordis, a system whose design is detailed in "A Programming Paradigm for Spatiotemporal Composability." Users should anticipate rapid iteration and compatibility-breaking changes during this preview phase. A Discord community is available for engagement.

Graft significantly enhances Claude Code, Cursor, Codex, and Gemini by improving correctness and efficiency. It achieved a 12-point increase in correctness, resolving 33 out of 50 instances compared to Cold Claude Code's 27. This improvement came with 23% fewer tokens, 25% fewer tool calls, and 32% less wall-clock time, leading to 19% cost savings. Graft also offers compiler-grade edges via `graft build --lsp` for languages like Rust, C/C++, Go, Python, and TS/JS, using language servers such as rust-analyzer and clangd.

Unsloth introduces Qwen3.8-27B GGUF files, featuring Unsloth Dynamic V3.0 for SOTA quantization. This new generation of the Qwen open-model family, built on Qwen3.5, offers significant improvements in coding, professional work, research, and long-horizon agentic tasks. Qwen3.8-27B is a compact, deployment-friendly 27B parameter model with native vision-language understanding, flexible thinking control, and enhanced agent execution for reliable multi-step task completion. It also includes developer role support and MTP for fast inference.

DeepSeek has launched DeepSeek Harness in developer preview under the MIT license. This new offering emphasizes a design where "everything is a plugin" that can be easily swapped out. DeepSeek is expanding its focus beyond just the model layer, aiming to provide software tools that developers can use to deploy AI agents effectively.

Google has introduced a new toggle in Gemini and Flow, allowing users to remove visible watermarks from AI-generated images, videos, and music. While the visible watermarks can now be removed, Google will continue to embed invisible SynthID watermarks and C2PA metadata into these AI-generated media files. This update provides users with more control over the appearance of their AI-generated content while still maintaining a level of traceability through embedded metadata.

Alibaba has released the weights for its Qwen3.8 models under an Apache 2.0 license. This release includes the Qwen3.8-27B model, which Alibaba states surpasses its predecessor, Qwen3.7-Plus, in overall performance. The Qwen3.8-27B is described as a native multimodal dense model with 27 billion parameters, excelling particularly in real-world coding and office workflows. It also features a 262K native context, extendable to 1M.

Cerebras and OpenAI have introduced Ultrafast Mode, a new service tier for the OpenAI API, powered by Cerebras. This mode, initially available to select customers, accelerates GPT-5.6 Sol to deliver up to 750 output tokens per second without compromising quality. In evaluations, GPT-5.6 Sol on Ultrafast mode answered 2,500 HLE questions in 11 hours and 11 minutes, nearly 7 times faster than Claude Fable 5, which took 78 hours and 27 minutes for the same task.

Greg Brockman is reportedly in "founder mode," increasing his involvement across all levels of OpenAI. This move aims to build out a leadership team in anticipation of an expected IPO. This development follows a wave of departures from top OpenAI executives, including Sam Altman's top deputy, a longtime chief operating officer, and the chief revenue officer.

Future Claude models will incorporate text watermarking to indicate the likelihood of AI involvement in text generation, aligning with the EU AI Act. This watermark subtly influences token choices, but is less applied to exact content like code or simple sums. Additionally, Claude will attach C2PA content credentials, a cryptographically signed note in metadata, to supported file types like .png, .jpg, or .svg, to show that the file was made or processed with Claude.

Mole is a deep-research agent designed for terminals, featuring an enforced budget and a privacy boundary for local data. It boasts a 0% budget overshoot, 100% claim integrity with verified quotes and sources, and 100% citation accuracy. The grounding rate is 80%, with a precision of 70% with the confirm pass (51% without). Merge precision/recall stands at 1.000/1.000 on constructed ground truth. Mole is released under the Apache-2.0 license.

Anthropic has updated its misalignment risk estimate from "very low" to "low," according to a risk report. The company also stated that it does not intend to release an internal model known as "Model 2." This decision comes despite "Model 2" appearing to be more powerful than their current top-tier model, Mythos. The report, covered by Madison Mills for Axios, highlights Anthropic's cautious approach to deploying advanced AI models.

Amy Kremer, a pro-Trump activist, played a role in organizing the January 6 rally and now chairs Humans First, a grassroots anti-data center group. Her strong loyalty to Trump is seen as a key factor in mobilizing support for this new cause, as detailed in a profile by Veronica Irwin for Transformer.