Qwen3.6-35B-A3B opens 35B MoE weights with 3B active parameters
Original: ⚡ Meet Qwen3.6-35B-A3B：Now Open-Source！🚀🚀 A sparse MoE model, 35B total params, 3B active. Apache 2.0 license. 🔥 Agentic coding on par with models 10x its active size 📷 Strong multimodal perception and reasoning ability 🧠 Multimodal thinking + non-thinking modes Efficient. Powerful. Versatile. Try it now👇 Blog：https://qwen.ai/blog?id=qwen3.6-35b-a3b Qwen Studio：https://chat.qwen.ai HuggingFace：https://huggingface.co/Qwen/Qwen3.6-35B-A3B ModelScope：https://modelscope.cn/models/Qwen/Qwen3.6-35B-A3B API（‘Qwen3.6-Flash’ on Model Studio）：Coming soon～ Stay tuned View original →
What the tweet revealed
The Qwen account wrote “Meet Qwen3.6-35B-A3B: Now Open-Source” and described the model as a sparse MoE with 35B total parameters, 3B active, and an Apache 2.0 license. That is high-signal because Alibaba is not only shipping an API endpoint; it is releasing open weights for a model meant to cover agentic coding, multimodal perception, and reasoning modes.
Alibaba_Qwen is the official model account for Qwen releases, and the post was amplified by Hugging Face’s feed, which is why it surfaced strongly in the crawl. The linked Qwen blog and Hugging Face card give the technical substance behind the tweet: Qwen3.6-35B-A3B is a fully open-source MoE model, available through Qwen Studio and open-weight hosting, with API access for the related Qwen3.6-Flash path coming through Model Studio.
The benchmark signal
The strongest coding numbers are in the blog’s agentic table. Qwen lists 73.4 on SWE-bench Verified, 67.2 on SWE-bench Multilingual, and 51.5 on Terminal-Bench 2.0. Those scores are the reason the tweet’s claim about coding “on par with models 10x its active size” matters: the model only activates 3B parameters at inference time, so the practical question is whether it can keep costs low while preserving agent behavior.
The model also targets multimodal use. Qwen highlights strong perception and reasoning, plus thinking and non-thinking modes. That mix puts it in the same competitive lane as compact agent models that need to inspect screenshots, reason over code, and call tools without demanding frontier-model serving budgets. Open weights and Apache 2.0 licensing make it easier for labs and companies to test those claims inside their own stacks.
What to watch next is whether independent SWE-bench and Terminal-Bench runs reproduce the official numbers, how inference cost compares with dense models, and whether Qwen3.6-Flash API availability closes the gap between open-weight experimentation and managed deployment. Source: Qwen X post · Qwen blog · Hugging Face model page
Related Articles
Google said on March 26, 2026 that Search Live is expanding to every language and country where AI Mode is already available. The rollout reaches more than 200 countries and territories and uses Gemini 3.1 Flash Live to make search more conversational, voice-first, and camera-aware.
Google expanded Search Live on March 26, 2026 to every language and location where AI Mode is available. The move pushes multimodal voice-and-camera search to more than 200 countries and territories and gives Gemini’s live audio stack a much larger real-world footprint.
Mistral promoted Voxtral TTS on X on March 26, 2026. Mistral's release post describes a 4B-parameter multilingual TTS model with nine-language support, low time-to-first-audio, availability in Mistral Studio and API, open weights on Hugging Face under CC BY-NC 4.0, and pricing at $0.016 per 1,000 characters.
Comments (0)
No comments yet. Be the first to comment!