Alibaba Classifies Claude Code as "High-Risk" and Bans Internal Use

According to a TechCrunch report, Alibaba has banned the use of Anthropic's CLI coding tool "Claude Code" within the company. Claude Code has been classified as "high-risk software," and the restriction is believed to be part of the company's internal security policy.

This move signals that, for Chinese tech companies, sourcing AI development tools is no longer just a matter of technology selection but is becoming a geopolitical and compliance issue. Meanwhile, the community is actively comparing Claude Code against alternative tools (such as Hermes), and there are reports that running StepFun's Step 3.7 Flash model on Claude Code produced better results than going through Hermes.

DeepSeek V4 Local Adoption Accelerates — Quantized KV Cache Fix and Ecosystem Support

Quantized KV Cache Fix Merged into llama.cpp

A post on Reddit's LocalLLaMA community reported that quantized KV cache fixes have been integrated for the DeepSeek V4 branch. A branch merging three llama.cpp PRs (#25247, #25303, and #25202) has been made public, with perplexity values measured across f16, Q8_0, and Q4_0 configurations.

The results show that even with Q4_0 quantization, perplexity came in at 4.05 (+0.03 versus f16), keeping accuracy loss to a minimum and confirming that practical performance can be maintained even in environments with tight VRAM constraints. The DeepSeek V4-Flash model has been tested on Blackwell-generation GPUs, with operational validation in high-end environments moving forward.

Compatibility Testing with Codex and Claude Code

An article published on Qiita examined whether DeepSeek V4 Pro can be used as an endpoint for OpenAI Codex and Claude Code. Attempts to use local LLMs as backends for coding assistants are on the rise, and compatibility with the major tools is a key practical consideration.

China's UBTech Unveils AI-Powered Humanoid Robot

According to Reuters, China's UBTech has unveiled a realistic humanoid robot with built-in AI. The company has long focused on the humanoid robotics field, and this model is positioned as a companion robot with integrated AI capabilities.

Backed by government support, Chinese robotics companies are notably pulling ahead in the humanoid space. It is a case worth watching as an example of advancing AI models combining with physical interfaces (robots).

Where Mistral AI Stands — The French Startup Taking on OpenAI

A TechCrunch explainer rounds up where Mistral AI stands today. Founded in 2023, the company has raised substantial funding and, under its vision of putting "frontier AI in everyone's hands," has established itself as a European AI company rivaling OpenAI while also offering some open-source models.

Mistral's approach, which runs on the twin tracks of an open-weights strategy and a commercial API, is easiest to understand when viewed in the context of comparisons with other players such as DeepSeek and Qwen.

Local Model Benchmark: Qwen3.6-27B Closes In on Larger Models

Reddit's LocalLLaMA saw benchmark results shared for eight local models, covering fantasy RP and agent-style tasks. The benchmark uses an external LLM as the judge, evaluating metrics such as quest completion rate, scene termination, item tracking, character detection, and narrative structure.

On overall score, Gemma-4-31B took the top spot at 87%, followed by Qwen3.6-27B at 82% — a strong showing considering its parameter count. What's interesting is that even models with similar overall scores show large per-category strengths and weaknesses, with some models noted to drop sharply on "NPC reasoning" and "quest summarization."

---