If you've been paying attention to open-source LLMs lately, you can't have missed Qwen.
Alibaba's Qwen team has been iterating at a breakneck pace: Qwen 3.5 just settled in, then 3.6 dropped with flagship 27B coding ability, and before anyone could catch their breath, Qwen 3.7 Preview was out.
The discussion heat on Reddit and Hacker News tells the whole story.
Qwen 3.6-27B: Flagship Coding in 27B Parameters
This article scored 993 upvotes and 458 comments on HN — one of the most discussed open-source model announcements recently.
Title says it all: "Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model." The core message: at 27B parameters, Qwen 3.6's coding ability competes with models several times its size.
For local deployment, this is huge. 27B is near the upper limit of what consumer GPUs (24GB VRAM) can run for a dense model. Qwen 3.6 achieving flagship-level coding at this size means you can run a coding assistant on a single 4090 or 3090 that doesn't compromise on capability.
Qwen 3.6-Max-Preview: Bigger Ambitions
The same day also saw Qwen 3.6-Max-Preview (705 upvotes). This version aims higher — directly targeting GPT-class closed-source models. It's just a preview, but early community feedback is strong, especially on reasoning ability and multi-turn conversation.
Qwen 3.7 Preview: Fresh Off the Press
By Qwen 3.7 Preview, the pace had clearly accelerated. The gap between 3.5 and 3.7 keeps shrinking, showing Alibaba's training pipeline is running smoothly. Each minor version update brings perceptible improvements — no牙膏-style incremental upgrades here.
Orthrus: 7.8x Throughput Boost
Another project worth noting is Orthrus-Qwen3 (243 upvotes on HN). It optimizes Qwen 3 inference to achieve up to 7.8x token/forward throughput while maintaining identical output distribution to the original model.
This shows the Qwen 3 ecosystem isn't just the official model — community optimization tools are keeping pace too.
Some controversy
It's not all praise. One article (60 upvotes) analyzed the political censorship mechanisms inside Qwen 3.5's weights, revealing which topics trigger restricted responses. It's a valid concern for an open-source model — the weights may be open, but training data biases remain baked in.
My take
Qwen's current iteration pace reminds me of the Llama series' rapid evolution in 2023-2024. But Qwen feels different: their training seems more purposeful. Each update targets a specific direction — coding ability, reasoning capability, MTP support. These aren't "just trained on more data" releases. Architectural evolution is happening in parallel.
If this pace continues, I wouldn't be surprised to see Qwen become one of the de facto standards in open-source LLMs.
