Google releases HEIR homomorphic encryption compiler for private AI

01.Google Ships HEIR, a Compiler That Lets Anyone Run AI Over Encrypted Data

Google published HEIR (Homomorphic Encryption Intermediate Representation), an open-source compiler toolchain in a security blog post that converts pre-trained AI models to operate on encrypted data without ever decrypting them. The post comes from Jeremy Kun of the Google security team and marks a milestone for private AI inference, where the old trade-off has always been stark: end-to-end encryption keeps data safe but blinds the service provider from using it for anything useful, while local processing on user devices runs up against limited compute power or risks leaking proprietary models.

Homomorphic encryption resolves this by running computations directly on ciphertexts and returning encrypted results. Servers process encrypted inputs without ever touching the plaintext. A cloud recommendation system can recommend content without seeing which videos a user has watched. That capability exists already in Google's own demo, and the post shows it working across four real applications built on HEIR: private content recommendations jointly developed with Belfort Labs and New York University, credit card fraud detection compiled alongside Niobium and hardshell.ai, encrypted network traffic anomaly detection using the Kitsune system, and a hotword detector that lets audio-triggered AI agents recognize wake words without exposing recordings.

The toolchain has accumulated partnerships with hardware accelerators from Belfort Labs, Niobium, Cornami, and Optalysys, plus academic collaborations with Georgia Tech, Carnegie Mellon, UC Santa Barbara, Purdue, the University of Edinburgh, Tsinghua, and several others. Four peer-reviewed publications have been built on HEIR since Google announced its intentions in 2023, with more in preparation. The vision is straightforward: make encrypted inference a one-click task for developers who are not cryptographers.

thinkidiot take: Homomorphic encryption has existed for decades but always stayed academic because the cost overhead was prohibitive and building with it required specialist teams. Google's HEIR shifts both needles at once, which would normally give me reason to be skeptical of press claims. The real test will be whether practitioners actually adopt it for production workloads where latency tolerances are tight.

02.OpenAI Launches Ultrafast Mode: GPT-5.6 Sol at 750 Tokens Per Second

OpenAI released a preview of its Ultrafast inference mode at up to 750 output tokens per second, powered by a Cerebras hardware partnership announced earlier this year under a ten-billion-dollar deal. The service delivers up to 750 output tokens per second from the flagship model GPT-5.6 Sol, which is described as up to 14 times faster than standard serving.

The acceleration comes entirely from Cerebras' wafer-scale infrastructure rather than any architectural changes in the model itself. Ultrafast will initially be available only through the OpenAI API for GPT-5.6 Sol and limited to select API customers before a broader rollout. This represents an opening salvo in what is effectively a speed war between inference providers, with Google's HEIR showing that at least some of them are thinking about how fast you can run private models too.

thinkidiot take: The Cerebras partnership shows OpenAI betting big on specialized silicon over more GPUs. A 750-token-per-second output rate is impressive for batch inference, but whether it matters for individual chat sessions depends on the round-trip latency of the network and rendering layer, not just model throughput. Speed wars always get measured where it's most convenient, not where users actually care.

03.Alibaba Qwen Releases Qwen 3.8: A Dense Model That Outperforms Larger Predecessors

Alibaba's AI team Qwen released open-weight versions of Qwen 3.8 under the Apache 2.0 license with performance claims. The core model, Qwen3.8-27B, is a multimodal dense model with 27 billion parameters. According to Qwen, it outperforms the larger Qwen3.7-Plus in coding and office tasks despite having fewer parameters, which represents a genuine efficiency gain rather than another brute-force scale-up.

The model also shows improved agent capabilities: it can plan more independently and complete multi-step tasks more reliably than its predecessor. It natively handles up to 262,000 tokens of context, which covers most practical document analysis and codebase navigation workloads without external chunking strategies. For teams constrained by licensing due to compliance requirements, Apache 2.0 is permissive enough for commercial use with no copyleft obligations.

thinkidiot take: The open-weight space around coding models has been crowded for months, but a denser model beating a larger one on benchmarks matters because it means deployment gets cheaper and easier. Whether the improvement holds up on real-world codebases is another question entirely, but the Apache 2.0 license makes this easy to test in your own environments.

04.Zhipu AI Releases GLM-5.3: Strongest Open-Weights Coding Model, Claims Its Creators

Chinese startup Zhipu AI released GLM-5.3, described by the team as the most powerful open-weights coding model currently available. The model shares its base weights with GLM-5.2; all gains come from extended post-training on new data and task environments rather than architectural changes.

The biggest measured improvements appear in agent-based tasks, where Zhipu says they built custom training data including cybersecurity environments that represent a gap Chinese models typically face compared to US frontier models like Kimi or Qwen. The claim here is not just raw coding performance but autonomous problem-solving capability over extended interaction chains.

Zhipu's track record of incremental GLM releases has been steady rather than revolutionary, making this one worth watching because it indicates whether the 27-billion-parameter class of open-weight models is closing on frontier capabilities through better training methodology alone.

thinkidiot take: When every Chinese model claims to be "the strongest open-weights coding model," you should assume the claim is at least optimistic. Post-training-only gains of this magnitude would be remarkable if verified. This is editorial selection of an interesting data point, not a judgment that one Chinese model beats another. Verify the benchmarks yourself.

ML Papers Worth Reading

The following papers were published or updated in the previous 24 hours and selected for practical relevance to working technologists. These are editorial picks, not an objective ranking, and have not been peer reviewed.

• Vero: Can AI Agents Build Formally Verified Software Repositories? (cs.LG, cs.AI) - Tests whether AI agents can produce both implementations and machine-checked proofs simultaneously. Verifiable code generation would change the risk calculation on any automation that touches production systems. https://arxiv.org/abs/2608.13522v1

• OmniScientist: An Omni-Modal, Omni-Discipline AI Scientist (cs.AI, cs.CL) - Demonstrates foundation models automating workflows across hypothesis generation through manuscript preparation with access to tools and domains previously handled by human researchers. https://arxiv.org/abs/2608.13558v1

• SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization (cs.CL) - Addresses a real problem in interpretability research: sparse autoencoders are useful for extracting features from language models but explaining what those features mean has always relied on superficial external observation rather than grounded feature-level analysis. https://arxiv.org/abs/2608.13538v1

• Intern-S2-Preview: Scientific Agentic Foundation Model (cs.LG, cs.CL, cs.CV) - Presents a foundation model that handles heterogeneous scientific evidence over multiple modalities with sustained multi-turn reasoning across long task horizons, relevant to anyone who needs AI to read the room and keep working through complex domains. https://arxiv.org/abs/2608.13505v1

• Synthetic Persona Pretraining: Alignment from Token Zero (cs.LG, cs.AI, cs.CL) - Proposes that alignment and assistant identity itself should be trained rather than emergent, using synthetic persona data to align model goals with human values from the very start of training. https://arxiv.org/abs/2608.13482v1

AI Repository of the Day

Today's repository is cactus-compute/needle, a 14-megabyte foundation model designed for deployment on tiny devices: phones, wearables, smart home hardware, and robots. At that size, Needle runs entirely on-device with no cloud dependency, which matters for privacy-sensitive applications or environments with unreliable connectivity. It is aimed at embedded systems developers and anyone who has tried to get a useful model onto resource-constrained hardware and hit the wall of model file sizes.