Your evening AI news roundup for November 3. The headline story is China's shifting AI landscape. According to an industry paper report, DeepSeek's next-generation model "V5" is planned to be pretrained on Huawei's AI chip, the Ascend. With US export controls still in effect, if domestic chip adoption expands from "inference" to "training," China's AI compute self-sufficiency will reach a major milestone. Elsewhere on the China side, IDs of upcoming Qwen models leaked in a GitHub repository. On the US side, Palantir's earnings outlook flashed a growth slowdown signal, while OpenAI has announced a bug bounty increase and security enhancements in quick succession.
DeepSeek V5 Plans Pretraining on Huawei Ascend — From Inference to Training
According to a report by a reporter for the Chinese industry paper STAR Market Daily, citing information obtained from multiple industry supply chain sources, DeepSeek's next-generation large-scale model, "DeepSeek V5," is planned to be pretrained using Huawei's Ascend chips. Until now, Ascend chips have mainly been used for inference with DeepSeek's large models.
The two companies have a history of collaboration. During the 2025 Spring Festival (Chinese New Year), Huawei Cloud, acting on instructions from Huawei's top leadership, reportedly deployed more than 3,000 Ascend chips at its Gui'an data center and completed the deployment and verification of DeepSeek-V3/R1 within days (as described in "The Rise of Ascend," a book by Fang Xingdong of Zhejiang University).
Why it matters: With US semiconductor export controls ongoing, securing compute resources has become the biggest constraint on China's AI development. If the shift in supply from NVIDIA chips to domestic Ascend extends beyond inference all the way to pretraining, it could shake the very effectiveness of the restrictions. Note, however, that the report describes this only as a "plan"—the scale and timing of the training run have not been officially confirmed.
Qwen4-max and qwen4-omni-120b Leak in GitHub Repository — 262k Context
Reddit users reported that IDs for unreleased models have appeared in the GitHub repository of "Qwen Code," Alibaba Qwen's official CLI. Benchmark records in pull requests include IDs such as "qwen4-max," "qwen4-plus," "qwen4.5-max," "qwen4-coder-plus," "qwen4-max-2026-10-01," and "qwen4-omni-120b-think," all of which reportedly resolve to a context length of 262,144 tokens. Notably, qwen4-omni-120b-think is said to already be used as the session model in test fixtures, suggesting it is a real model rather than a hypothetical ID.
This follows a recent official statement from a Qwen lead that "future Qwen 4.0 and Qwen 5.0 will remain open source." The community now sees this as a follow-up, with some observing that "a successor to Qwen3.8 Flash Next is close." Keep in mind, though, that the evidence is limited to traces in the repository—there has been no official announcement yet.
Palantir's Q4 Outlook Beats Analyst Estimates, but Growth Is Set to Slow
According to Bloomberg, Palantir Technologies, the AI data analytics company whose major clients include the US military, guided December-quarter (fourth-quarter) revenue to about $2.65 billion, beating the analyst consensus of $2.46 billion. Strong commercial sales are cited as the backdrop.
But the growth rate is what deserves attention. The guidance implies only about 15% growth over the prior period, well below the roughly 20% enjoyed in the most recent quarter. That would be the slowest growth on this metric since the second quarter of 2025—the first slowdown signal for a company that has grown rapidly on the back of AI expectations.
OpenAI Raises Bug Bounty Payouts — Security Sprint with Cathedral Systems Also in the Works
OpenAI announced that it is increasing payouts for its security bug bounty. Aimed at independent security researchers, the move introduces new bonuses for "report quality," "loyalty" (sustained contributions), and "priority campaigns."
The company also revealed that it is preparing its next internal security sprint in partnership with Cathedral Systems. The initiative uses AI to actively test and strengthen the company's own defenses—a sign that OpenAI is building out a "defended by AI" posture.
Apple Watch's Conversational Rewind AI Enters Beta Testing
A new Audio Intelligence feature for Apple Watch has entered beta testing. Its main capabilities are "Rewind," which lets you replay a conversation to check what the other person said, and "Recap," which summarizes the key points of a conversation. It is said to be aimed at use cases like "recalling what someone said" or "taking notes on a conversation." How far voice AI use cases unique to wearables can go in practice will be worth watching.
"Weight Space Programming": LLMs Directly Read and Write Other Models' Weights
On the research front, attention is focused on a finding highlighted by the AI newsletter Latent Space as "Weight Space Programming": having a sufficiently capable model directly "read and write" the neural network weights of a smaller model. Conventional autonomous model training has always involved human ML engineering—writing PyTorch/JAX code, reading loss curves, and optimizing kernels and data based on evaluation results. This work suggests those steps could be replaced by having LLMs operate directly at the weight level.
The cited examples include: (1) determining a model's behavior from its weights alone; (2) fixing a bug in one shot from a failure trace; (3) improving in a specific direction with around 50 steps of instruction; and (4) writing weights from scratch that can sometimes beat independently trained models. The effort has been likened to reading the 1s and 0s of a compiled executable directly, and independent replications are reportedly starting to appear. Note, however, that the targets are still very small models, and some voices urge caution against overstating the results.