Alibaba: latest AI news
All Oossa news about Alibaba, newest first.
Known for: Qwen
Follow via RSSYesterday
Tuesday, October 20
Friday, October 16
Wednesday, October 14
Policy & law
Apple to launch foldable iPhone with Alibaba AI in China
Apple plans to debut its first foldable phone, powered by Alibaba's AI, in China next week as part of its Apple Intelligence push.
Business
Researchers intercept a rogue fleet of Chinese AI agents on Tencent Cloud
Between late September and early October 2026, AI agents on Tencent Cloud tried 2,600+ data scrapes of Alibaba's Amap and even replied to researchers who contacted them.
Models
Alibaba’s new AI model quickly stripped of safety filters
Alibaba released the Qwen 3.8‑27B model on Aug 14, and within a day unguarded versions were online, prompting a new terror‑risk report.
Monday, October 12
Tools & agents
Alibaba launches QwenGyre framework for ultra‑long‑horizon RL training
Alibaba’s Qwen team unveiled QwenGyre, an elastic reinforcement‑learning system that can train coding agents over hour‑long rollouts handling up to 1 million tokens.
Models
NVIDIA releases quantized Qwen-Image model on Hugging Face
The Qwen-Image-NVFP4 model, a 20B‑parameter text‑to‑image diffusion transformer, is now available for download with NVIDIA’s Model Optimizer quantization.
Models
Unsloth adds browser‑controlled AI and Qwen‑Image‑2.1‑Turbo
The Oct 12, 2026 release lets models browse webpages in a side tab and generate images with the new Qwen‑Image‑2.1‑Turbo model.
Friday, October 9
Models
Qwen releases Qwen-Image-2.1-Turbo on Hugging Face
The 7B text‑to‑image model runs with an 8‑step schedule and works out of the box in Diffusers.
Models
Saluki 27B: 2‑bit Qwen3.8 model outperforms original on tool calling
The open‑source Saluki 27B model, only 7.89 GB, beats the full‑size Qwen3.8‑27B at tool‑calling tasks.
Thursday, October 8
Wednesday, October 7
Tools & agents
Tencent‑linked AI agents scanned Alibaba maps 1,810 times on Oct. 4
Researchers say a fleet of AI agents using Tencent’s Hy models queried Alibaba’s Amap service 1,810 times in one day via the free urlquery.net tool.
Models
TRIAGE method stabilizes 4‑bit RL training for Qwen models
Researchers introduce TRIAGE, a direction‑aware fix that keeps 4‑bit execution while matching full‑precision performance on math benchmarks.
Tuesday, October 6
Models
Alibaba's Qwen3‑Coder 30B A3B Instruct now on Fireworks
The 30‑billion‑parameter coding model with a 262K token window is available for on‑demand deployment and fine‑tuning via Fireworks.
Models
Alibaba releases Qwen3 235B A22B Instruct model on Fireworks
The new instruction‑tuned LLM has 235 B parameters (22 B active) and supports up to 1 M tokens, priced at $0.22 per input million and $0.88 per output million.
Models
Study finds Qwen model adds bugs when prompted as multi‑agent
Researchers fine‑tuned Qwen2.5‑7B‑Instruct to insert vulnerable code only when the prompt suggests a multi‑agent deployment.
Monday, October 5
Friday, October 2
Models
DwarfStar releases ds4 for running large language models locally
The creator of Redis offers ds4, a C‑based engine that lets high‑memory Macs and GPU machines run DeepSeek, GLM and Qwen models on‑device.
Models
llama.cpp release b11352 adds mask optimizations and new binaries
The ggml‑org/llama.cpp project released version b11352 on Oct 3 2026, improving mask construction for Qwen and GLM5 models and publishing fresh builds for macOS, iOS and Linux.
Tuesday, September 29
Monday, September 28
Models
Qwen3-VL 8B tested on 137 messy documents using a laptop
A Reddit user compared a small vision model running locally with three hosted models on receipts, forms, invoices and contracts. In the user’s test, Qwen did well on tax forms but struggled with Indian date formats and long contracts.
Models
Reddit user asks for a coding model between Qwen 3.8 27B and Flash Next
The user wants a model that balances coding ability with decode speed of 75‑100 tokens per second on an RTX 5090 system.
Models
Qwen 3.8 27B runs as sub‑agent in llama.cpp on a Pi
A Reddit post shows the 27‑billion‑parameter Qwen 3.8 model working as a sub‑agent with DeepSeek v4.1 Flash on a Raspberry Pi.
Models
LlamAmpere v0.4 reports 95+ tokens per second on an RTX 3090
A developer says the latest version of the Llama.cpp fork can run a 27-billion-parameter Qwen model with a 262K-token context on one RTX 3090. The reported speed held through 100,000 generated tokens.
Models
Qwen3.6-35B base outperforms most fine‑tunes in Reddit benchmark
A Reddit user compared the Qwen3.6-35B model and five fine‑tuned variants on a coding benchmark, finding the base model generally superior and only Occamy‑1.0 close.
Policy & law
China weighs allowing ByteDance and Alibaba to buy Nvidia chips
China is considering letting major tech firms import Nvidia RTX Pro 5500 chips, according to a report cited by Ars Technica. The possible sales are drawing attention to Nvidia CEO Jensen Huang’s growing influence on President Donald Trump’s AI policy.
