Qwen

“Take your blinkers off and look at other providers apart from qwen. I can count 4-5 models released in the last 2 weeks alone that are sub 9b models . And this is with out looking. There has probably been way more. Some are good. And some are mid.Your hilarious. Qwen just released the best 27b model for local use on consumer hardware. FOR FREE….! and your giving out that you can’t run it because you don’t have the hardware. A used 3090 is not classed as enterprise tier hardware.”

“Well, this was predictable. After Qwen fired Junyang Lin, the next models are no longer open source. Ignoring the small models for a minute, they’ve fully locked down all the big models. No Deepseek/GLM competitor. And all the rumors on chinese weibo now say that the small model Qwen team is gone, and that Qwen 3.6 (and maybe 3.7) was the last model Junyang Lin worked on. There's not going to be any open source small models from Qwen anymore. Labs that have released open source models more recen”

“I have read some people about it but today i made an account since my claude again hit limits after a simple question and like 4 lines of code. I was skeptical since my only experience with qwen were the small local models 35b and 27b. But it blew me away, i asked it to first gather and read the documentation of both apps. What i wanted it to make was a plugin for the one app a media database that would bridge it to the other a media player. It read in total about 50 pages and then gave a respon”

“I have a love hate relationship with Qwen now. I do not like how they are no longer offering “practical user end” LLMs anymore or open weights for their latest image generations. But I am impressed with their image edit. 2.0 was impressive enough but I’m blown away by how intricate 3.0 is that it can do very specific details with immense accuracy where it looks completely natural.”

“I've used this model since launch, including the preview, and burned through 1B tokens on the standard plan with an extra usage top up. Now working during night discount only with pro plan , my usage runs about 350–500M tokens per week till it runs out (depending on cache rate - I am using it with zcode which maintain 95% above most of the time), projecting to roughly 1.67B tokens/month with good caching, still less than I used on Codex's €100 plan, where I never even hit 50%. The model itself i”

“I just subscribed to Qwen Cloud’s Standard Token Plan, mainly to use it with an AI coding agent (Hermes), and after actually using it for a few days, I honestly don’t understand how this plan is supposed to be considered good value. The Standard plan gives you 10,000 Credits per week. Sounds reasonable, right? Until you actually use it. I burned through roughly 70% of my weekly Credits in only 3 days while running a normal agent workflow. I’m not running hundreds of agents, doing massive batch i”

“For the current prices its a great deal. I have the standard plan. left it overnight (discount time) running - used 125 million tokens (85% cache hit) and the weekly usage went up by 7 % only. Speed is the best part. Alibaba has the infra to avoid throttling. Quality wise Opus5>GPT5.6 SOl> Qwen 3.8 max = Kimi K3. Kimi is fine but for the same tasks thinks way too long and the improvements over qwen aren't that noticeable. I mainly work with backend and Deep learning related tasks though.”

“I am using it. I bought on release day Sunday. Night time discount is only on qwen 3.8 max preview . As a matter of fact 10x usage promotion is also only for qwen 3.8. I used 40M token during night, it cost me around 6% of weekly usage. I used 80M tokens today during day and it used around 40% of weekly limit. So, either their night time discount is more than 20% or they changed usage limits behind the scenes. Night time 20% off should be after 90% normal discount(10x promotion), so final cost s”

“TL;DR: Built AI agents and RAG systems for companies in pharma, banking, and legal over 6 months. Sharing details on domain-specific fine-tuning approaches, how I handled reasoning loops and medical acronym disambiguation, my approach to context management at scale, and what actually works in production. No standard benchmarks exist for this stuff - had to work with domain experts to evaluate entire agent workflows. 4-bit quantization works great, needed 6-12x H100s for 60+ concurrent users. Her”

“How is this so bad? I gave the app some documents to analyze for me only for it to misattribute, mix the chapters and events with previously uploaded documents and give incomplete information. Sometimes, I could upload a file and it's saying it doesn't have it. How?”

“Personal experience, but currently the most biggest problem is in peak time cache miss . it some times lags and took too much time. This cause cache miss and this is causing costs. At off peak I hit 95% cache hit with pi coding agent and it become 75% at peak”

“For the current prices its a great deal. I have the standard plan. left it overnight (discount time) running - used 125 million tokens (85% cache hit) and the weekly usage went up by 7 % only. Speed is the best part. Alibaba has the infra to avoid throttling. Quality wise Opus5>GPT5.6 SOl> Qwen 3.8 max = Kimi K3. Kimi is fine but for the same tasks thinks way too long and the improvements over qwen aren't that noticeable. I mainly work with backend and Deep learning related tasks though.”

“I spent some time yesterday after work trying out the new qwen3.6-35b-a3b model, and at least for me it's the first time that I actually felt that a local model wasn't more of a pain to use than it was worth. I've been using LLMs in my personal/throwaway projects for a few months, for the kind of code that I don't feel any passion writing (most UI XML in Avalonia, embedded systems C++), and I used to have Sonet and Opus for free thanks to Github's student program but they cancelled that. I've be”

“While the technical advancements from the Qwen team continue to impress, the release of Qwen 3.8 27B highlights a widening disconnect between open-weight progress and accessible consumer hardware. Demanding roughly 24GB of VRAM to run smoothly at practical quantization levels puts it well out of reach for standard setups. With the market prices for high-VRAM GPUs and fast memory remaining excessively high, keeping up with these larger releases locally is becoming an expensive luxury rather than ”

“Yes exactly; the 2b 4b 9b sizes were perfect. IMO 27b is at the almost the max you can run on consumer hardware(think 32 gb of vram like the 5090). With a proper quant and speeds. You can of course push 32 gb of vram even more, etc.. Right now our best bet is a bonsai release, or hopefully for qwen to drop the smaller 3.8 models”

“Qwen Code comes with a coding agent free tier that feels like unlimited: if one cannot or does not want to pay 20$/mo it's an excellent alternative. I've consumed hundreds of millions of tokens building an app, never got rate limited or kicked out. It's a huge memory hog (I had to tweak the Linux memory management just to get it started). The UI feels much more lagging than Claude , but still better than Gemini which is almost unresponsive on my mid-low tier laptop. The dad jokes it makes during”