Hugging Face's co-founder and CEO are publicly backing Alibaba's Qwen3.6-27B as a dense, Apache 2.0-licensed model that delivers strong agentic coding and multimodal results without the footprint of much larger closed systems, with community reports noting practical speed advantages in local deployments.
Unpopular opinion: While everyone is so hyped about Fable, GPT5.6 and other huge and expensive models, I think the real hero of the last few months is *Qwen 27b*.
Our ML/AI engineering teams are have been very surprised how good/fast this open model is.
Early users highlight fast inference and repository-level reasoning that competes with expectations from frontier-scale models, though long-term stability across extended agent workflows remains to be tracked at scale.
The 27B dense model reportedly exceeds the older Qwen3.5-397B-A17B MoE variant on SWE-bench Verified and Terminal-Bench, yet exact head-to-head runs against current closed competitors are not fully detailed in available releases.
Many users praise Qwen 27B as a standout open model because it delivers strong real-world performance in local inference and often outperforms far more expensive alternatives.
No Digg Deeper questions have been answered for this story yet.
True
Unpopular opinion: While everyone is so hyped about Fable, GPT5.6 and other huge and expensive models, I think the real hero of the last few months is *Qwen 27b*.
Our ML/AI engineering teams are have been very surprised how good/fast this open model is.
@kskrygan I’m having trouble with that model. It doesn’t follow directions. I gave it a list of 6 issues to fix in some firmware. It ignored that and made 6 other unrelated changes. After a few different instances of that, I’m getting better results with Deepseek-V4-Flash on Dwarfstar4.
@kskrygan @grok is Qwen free and how good it is for Kotlin, KMP, CMP code?
@kskrygan I'm team Gemma, but only because I deal with multilingual workflows a lot. Otherwise yeah Qwen both this and the MoE models are really good.
@kskrygan What kind of hardware are they running the model on and how many concurrent users?
@kskrygan The same with MiniMax M2.7. 100+ tok/sec from their API, and good enough quality
@barscnasln @kskrygan You can run it on a MacBook Air 32GB
@barscnasln @kskrygan it runs in 24gb of vram at reasonable quants -- in ML world thats tablestakes
@h1kz0r @kskrygan Which size? And what languages?
@kskrygan Yes! I've been testing usage of qwen 30b since few days with Hermes agent on 4070TI and its WAYYY better than I was expecting. It can maintain long thinking, chain tool calls, search through docs, websearch..
I think people are underestimating power of harness engineering.
@Whos_Ur_Doggy @kskrygan For concrete examples: Wrong or bad quants, abliterated models feel smoother, fixing tool calling, many offer custom templates that try to fix things.
Then your stack at multiple levels, how is your agent augmenting context, vector search, lots of support from backend.
@kskrygan Fable is ridiculous. This will make so much working software.
@barscnasln @Love2Code @kskrygan Yeah you're right it's also incredibly early so it's pretty amazing that we can even do what we can locally on that budget
@kskrygan Exactly, no comparison.
@kskrygan That model is a miracle
@kskrygan Even Jetbrains is using local models? Damn, that’s nice.
@kskrygan Still can’t count though 🤷🏻♂️
@kskrygan Now if we just got a 3.7 27B
@kskrygan Qwen is one of the most underrated, but it outperforms most probably of the models which are 10X more expensive
@kskrygan try ornith 35b
