Its really weird to compare it to glm 4.5 base and deepseek v3 , instead you know comparing it to qwen or other similar sized / similar performance models.
They're comparing base models. There aren't very many large recent base models available. Qwen didn't release theirs. I think HY's chart is only missing Trinity Large.
This is, sadly, to be expected, I believe that based on current interperetation of EU's AI Act, releasing large model to EU in the open means that you are responsible for Risk Control, which means that when someone uses it for something crazy, Tencent may get into trouble.
First large model from Tencent, too bad they didn't put MiniMax in the comparison, it should be comparing to MiniMax and Qwen-397B. Hopefully it has an interesting personality or strength in certain areas. Gonna be a while before we see a gguf for this.
I was ready to dismiss it outright as benchmaxxed slop factory. I gave it a very short but fair shot and I'd say it's actually not that bad. It would be a good starting point to finetune by a good research team. Hence, well, probably why they chose to publish it with the ass puckering license...
Good breadth of knowledge. Not as smart as comparable models and disappointing at coding
Daemontatox@reddit
Its really weird to compare it to glm 4.5 base and deepseek v3 , instead you know comparing it to qwen or other similar sized / similar performance models.
Klutzy-Snow8016@reddit
They're comparing base models. There aren't very many large recent base models available. Qwen didn't release theirs. I think HY's chart is only missing Trinity Large.
spaceman_@reddit
Nice to see another open weights model. It should be noted this one has a pretty restrictive custom license:
Problems for normal people:
Problems for corpos:
Also: 299B is too rich for my blood
aanzeijar@reddit
Did Tencent care about licenses when training the model?
AssistBorn4589@reddit
This is, sadly, to be expected, I believe that based on current interperetation of EU's AI Act, releasing large model to EU in the open means that you are responsible for Risk Control, which means that when someone uses it for something crazy, Tencent may get into trouble.
ForsookComparison@reddit
Mistral needs to come to the US like huggingface did. The EU is making their stance clear that they'll die on a hill of regulations.
segmond@reddit
First large model from Tencent, too bad they didn't put MiniMax in the comparison, it should be comparing to MiniMax and Qwen-397B. Hopefully it has an interesting personality or strength in certain areas. Gonna be a while before we see a gguf for this.
MokoshHydro@reddit
I've tested it on programming tasks and it looks like Hy3 is below Qwen3.6-27B for development.
DefNattyBoii@reddit
I'd love to see some comparisons with q4 quants
Dany0@reddit
I was ready to dismiss it outright as benchmaxxed slop factory. I gave it a very short but fair shot and I'd say it's actually not that bad. It would be a good starting point to finetune by a good research team. Hence, well, probably why they chose to publish it with the ass puckering license...
Good breadth of knowledge. Not as smart as comparable models and disappointing at coding
Mickenfox@reddit
I've been trying it on OpenRouter and it seems pretty good.
DistanceSolar1449@reddit
It feels below Qwen 3.6 27b (or GPT-5.4 mini, which I’ve been using a lot of recently because it gives me more limits on a $20 ChatGPT plan)
Kodix@reddit
This seems kind of tiny (300B params) for its level of performance, no? Nice. Shame about the shitty license.
Ok_Technology_5962@reddit
Nice! More options! I like the 22b active sizeing
imp_12189@reddit
It's free on openrouter https://openrouter.ai/tencent/hy3-preview:free
Namra_7@reddit (OP)
base