April 16, 2026
The AI corpse that wouldn’t stay dead
Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All
Open-source underdog drops a free coding beast and the crowd goes wild
TLDR: Qwen just released a powerful free coding AI model that seems to rival much larger systems, shocking people who thought the project was crippled by staff exits. Commenters are split between cautious comparison to paid giants and loud celebration that real open, no-subscription AI isn’t dead yet.
The Qwen team just dropped a new free AI coding model and the comment section is treating it like a surprise comeback tour. After rumors that the project was doomed when key researchers walked out, one user summed up the collective shock: “Wowzers… that’s a huge drop. It’s better than 27B?” Translation: this smaller, cheaper model appears to punch above its weight.
The real drama? People thought Qwen had been “kneecapped” after internal shake-ups and high‑profile departures. Instead, they’ve come back with an open model that anyone can download, use, and tinker with. One commenter called it a relief just to see open weights still being released, turning the thread into a mini‑protest against closed, paywalled AI.
Others are already doing the math: how close is this to big paid models like GPT or Claude’s Sonnet, and is “free” good enough? The bargain-hunter energy is strong: if it’s even close to top commercial AIs, devs are ready to switch. Meanwhile, the hype-averse crowd is loving the zero‑BS vibe: no “too dangerous to release”, no IPO buzzwords, no subscriptions — just a direct link to download on Hugging Face. For now, the internet’s verdict: Qwen might be bruised, but it’s definitely not dead.
Key Points
- •Qwen released Qwen3.6-35B-A3B, a 35B-parameter mixture-of-experts model with 3B active parameters, as fully open-source with open weights.
- •The model targets strong agentic coding performance and multimodal perception/reasoning, outperforming its predecessor Qwen3.5-35B-A3B and rivaling larger dense models like Qwen3.5-27B and Gemma-31B on many benchmarks.
- •Qwen3.6-35B-A3B is accessible via interactive chat on Qwen Studio, will be available as Qwen3.6-Flash on Alibaba Cloud Model Studio API, and can be downloaded from Hugging Face and ModelScope.
- •Extensive benchmark evaluations across coding, agent, and tool-use suites (e.g., SWE-Bench variants, Terminal-Bench 2.0, Claw-Eval, SkillsBench, QwenWebBench, Tool Decathlon, MCPMark, MCP-Atlas) show competitive or superior performance to peer models at similar or larger scales.
- •The model also performs competitively on general knowledge, STEM, and reasoning benchmarks such as MMLU-Pro, MMLU-Redux, SuperGPQA, C-Eval, GPQA, LiveCodeBench v6, HMMT variants, IMOAnswerBench, and AIME26, highlighting a strong balance between efficiency and capability.