February 28, 2026
Sonnet at home or smoke and mirrors?
Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers
China’s ‘Sonnet at home’ AI drops—fans cheer, skeptics want receipts
TLDR: Alibaba’s new open models claim Sonnet 4.5-level power on home PCs with giant document handling. The crowd is split: fans hype fewer filters and strong tools, skeptics want real benchmarks and non‑Chinese options—fueling a louder local‑AI vs cloud trust showdown.
Alibaba just tossed a hand grenade into AI chatrooms: new open models called Qwen3.5 that they say can match Anthropic’s Claude Sonnet 4.5—on your own PC. You can grab them on Hugging Face or ModelScope, and the buzz is that they handle massive docs thanks to “quantization” (shrinking the model to run lighter), with some bragging about a million‑token memory on a beefy consumer graphics card. The vibes? Equal parts hype and side‑eye.
The comments are pure chaos. One-liners like “Yes.” turn into a full culture war: some cheer that Chinese models are “less lobotomized” than U.S. ones, while others demand hard proof. “Is there a non‑Chinese open rival?” asks one user, poking the geopolitics bear. Another fires back: show us real tests before we crown a new champ. Practical devs chime in: the new 35B model “is great” for tools but has quirks with Claude’s coding add‑ons. Meanwhile, the capybara PC image sparked memes about “rodent‑approved rigs” and “Sonnet at home” DIY energy. The big takeaway: if this holds up, local AI that rivals pricey cloud bots could be back in fashion—but until the evals land, it’s receipts or it didn’t happen. Read more via VentureBeat.
Key Points
- •Alibaba released the Qwen3.5 Medium Model series with four LLMs supporting agentic tool calling.
- •Three models (35B-A3B, 122B-A10B, 27B) are open source under Apache 2.0 and downloadable via Hugging Face and ModelScope.
- •Qwen3.5-Flash is proprietary and available through Alibaba Cloud’s Model Studio API, with noted cost advantages.
- •Qwen3.5 models reportedly outperform OpenAI’s GPT-5-mini and Anthropic’s Claude Sonnet 4.5 on third-party benchmarks.
- •Qwen3.5-35B-A3B can exceed a 1 million token context window on 32GB consumer GPUs via near-lossless 4-bit weight and KV cache quantization; architecture uses Gated Delta Networks with sparse MoE.