Qwen 3.6 Plus is a decent model in our benchmarks (which found it to perform lower than its model card) at gertlabs.com, but not ground-breaking.
The reason for the insane popularity is because it's pretty good AND free. It's a no-brainer to switch to this for anything usage-based that isn't frontier coding while the free limits are available. It's probably running a model ~100B parameters under the hood, which won't be so heavily subsidized for long.
EDIT: our tool usage benchmark is still running, but so far, its performance with tools is dramatically better than its one shot performance. I'm treating Qwen 3.6 Plus as a near-SOTA model now.
You will be rate limited, so it depends on your use case. We only ran into brief, intermittent short term rate limits when making thousands of calls for the benchmark, so I imagine it's fine for personal use.
If anyone wants to avoid twitter, just set up a browser redirector that sends you to the relevant xcancel page. I haven't seen the native Ttitter interface for months now. Don't put the expectation on HN when it's so easy to fix yourself.
I’m very curious if we’re going to ever get another “deepseek moment. Qwen is starting to feel like it could be one. But for it to be people would have to decide to care. It took about a month, I think mid December-mid January, from the deepseek paper for the “moment” so it doesn’t necessarily have to be right away.
What's gone unnoticed with the Gemma 4 release is that it crowned Qwen as the small model SOTA. So for the first time a Chinese lab holds the frontier in a model category. It is a minor DeepSeek model, because western labs have to catch up with Alibaba now.
depends on usage, Gemma 4 is better on visuals/html/css and language understanding (Which probably plays a role in prompting). But it's worse at code in general compared to Qwen 3.5 27B.
ANTHROPIC_BASE_URL=https://openrouter.ai/api ANTHROPIC_AUTH_TOKEN=<make_an_account_on_openrouter_and_get_this_from_the_settings_panel> claude --model qwen/qwen3.6-plus:free
> This repository has two ways of packaging Nix packages: defining them via pkgs/top-level/all-packages.nix (the old way); or defining them via the pkgs/by-name directory (the new way). Let's port my_example_package over to the new way.
i'm not actually working in the nixpkgs repo -- i'm trying these in a private repo that has very similar structure. i'm also a n00b with these tools, so probably a bad prompt. but Qwen 3.6 actually conflates "the old way" with "the new way", attempts to do the porting in reverse, and just gets stuck. gemma-4 E4B does better. even gpt-oss-120b, an open weight model from a _year_ ago, does the full port unattended.
so either it's shit at coding, or i'm using it wrong. curious to hear other anecdotes.
Qwen 3.6 Plus is a decent model in our benchmarks (which found it to perform lower than its model card) at gertlabs.com, but not ground-breaking.
The reason for the insane popularity is because it's pretty good AND free. It's a no-brainer to switch to this for anything usage-based that isn't frontier coding while the free limits are available. It's probably running a model ~100B parameters under the hood, which won't be so heavily subsidized for long.
EDIT: our tool usage benchmark is still running, but so far, its performance with tools is dramatically better than its one shot performance. I'm treating Qwen 3.6 Plus as a near-SOTA model now.
is it unlimited free, or the usual openrouter free (50 or 1000 requests/day)
You will be rate limited, so it depends on your use case. We only ran into brief, intermittent short term rate limits when making thousands of calls for the benchmark, so I imagine it's fine for personal use.
https://xcancel.com/openrouter/status/2040239467865489874
[flagged]
I prefer my links unrewritten
Replying to you since the rage guy is flagged.
If anyone wants to avoid twitter, just set up a browser redirector that sends you to the relevant xcancel page. I haven't seen the native Ttitter interface for months now. Don't put the expectation on HN when it's so easy to fix yourself.
I’m very curious if we’re going to ever get another “deepseek moment. Qwen is starting to feel like it could be one. But for it to be people would have to decide to care. It took about a month, I think mid December-mid January, from the deepseek paper for the “moment” so it doesn’t necessarily have to be right away.
What's gone unnoticed with the Gemma 4 release is that it crowned Qwen as the small model SOTA. So for the first time a Chinese lab holds the frontier in a model category. It is a minor DeepSeek model, because western labs have to catch up with Alibaba now.
on my 16 GB GPU Gemma 4 is better and faster than Qwen 3.5, both at 4-bit
so it's not so clear cut
depends on usage, Gemma 4 is better on visuals/html/css and language understanding (Which probably plays a role in prompting). But it's worse at code in general compared to Qwen 3.5 27B.
It's unnoticed because it didn't. In Google's own benchmarks they are on par, and I've seen 3rd party benchmarks where Qwen beats G4 with high margin
The day a western anything will need to catch up with alibaba will be a notable day indeed. Also, this will never happen.
I wonder what kind of workloads people are putting through it. Presumably all tokens submitted are used to train on.
If it overthinks everything the way Qwen 3.5 running locally does, then I am not surprised! :)
Anybody want to give an anecdotal take on how good it is?
sure.
git clone https://github.com/nixos/nixpkgs
ANTHROPIC_BASE_URL=https://openrouter.ai/api ANTHROPIC_AUTH_TOKEN=<make_an_account_on_openrouter_and_get_this_from_the_settings_panel> claude --model qwen/qwen3.6-plus:free
> This repository has two ways of packaging Nix packages: defining them via pkgs/top-level/all-packages.nix (the old way); or defining them via the pkgs/by-name directory (the new way). Let's port my_example_package over to the new way.
i'm not actually working in the nixpkgs repo -- i'm trying these in a private repo that has very similar structure. i'm also a n00b with these tools, so probably a bad prompt. but Qwen 3.6 actually conflates "the old way" with "the new way", attempts to do the porting in reverse, and just gets stuck. gemma-4 E4B does better. even gpt-oss-120b, an open weight model from a _year_ ago, does the full port unattended.
so either it's shit at coding, or i'm using it wrong. curious to hear other anecdotes.
How does OpenRouter manage to run closed-weight models like Qwen 3.6? Did Qwen have to actually cooperate with them by contributing the weights?
OpenRouter is primarily a router. It just proxies requests through to the actual provider.
gpt-oss-120b is vastly better than gemma-4 E4B