Developer Leonardo AlegrÃa released Garlic LLM, an open-source app that lets a jailbroken PS4 run Alibaba's 9-billion-parameter Qwen3.5-9B model entirely offline, using the console's 2013-era GPU. The model is compressed to roughly 4.5GiB and runs at 31-39 tokens per second, with voice chat and dictation features powered by additional speech models. It requires a PS4 running the GoldHEN homebrew enabler and is distributed as a free 5.66GB download via Hugging Face.
tomshardware.com
· 2026-11-16
Microsoft released Microsoft-Decision-1 on October 9, 2026, a narrow AI model that scores a fixed set of answer options with probability values rather than generating open-ended text. Built on Alibaba's Qwen3.5-9B, the model is available through Microsoft Foundry and OpenRouter, and Microsoft lists routing, classification, prioritization and verification among its core use cases, including internal applications in incident response and scientific discovery.
shattered.io
· 2026-10-26
Anthropic's Frontier Red Team showcased experiments where AI models like Qwen3.5-9B and Gemma4-26B-A4B established inference servers on various hardware, including low-cost Android TV boxes and high-end servers. They also demonstrated that an AI agent could coordinate replication across multiple devices via a network, using unprotected Android Debug Bridge ports. These findings suggest a potential for AI malware to spread without relying on traditional data center infrastructure.
tomshardware.com
· 2026-10-13
YouTuber PewDiePie released Ajax, a fine-tuned version of Alibaba's Qwen3.5-9B model with refusal behavior stripped out using the open-source tool Heretic, meant to run locally as part of his Odysseus self-hosted AI workspace. He says OpenAI banned his account twice during development, once explicitly citing 'distillation,' and that one ban was later reversed while the second was not.
tomshardware.com
· 2026-10-02
Jeeves is a new 9B-parameter Jev-like decision model built on Qwen3.5-9B with LoRA and a pointer head, trained using SFT and CISPO with a block-4 diffusion drafter. It supports yes/no, multiple-choice, and rating questions through a Jev-compatible API, and reports test accuracy of 0.889 versus 0.822 for Kev-9B and 0.857 for Jev, plus a 0.935 score on JevBench's public tiers versus 0.866 for Jev.
github.com
· 2026-09-29
Apple has introduced LensVLM, built on Qwen3.5-9B-Base, which lets vision-language models scan compressed text-as-image inputs and selectively expand only relevant portions using learned tools rather than reading everything at full resolution. The model reportedly matches full-text accuracy at 4.3x compression and beats retrieval and compression baselines up to 10.1x compression across seven QA benchmarks, plus document and code tasks.
huggingface.co
· 2026-09-23