Tencent @TencentHunyuan dropped the non-preview version of Hy3 and changed their license from the community one (restrictive + not allowed in SK, UK, EU) to Apache 2.0!!! 🙏
10:58 PM · Jul 5, 2026 · 23.5K Views
Tencent releases its 295B Hunyuan3 MoE model under an Apache 2.0 license, lifting regional access limits
The model matches larger systems on SWE-bench and GPQA Diamond.
575814314492.7K
Sentiment
Positive users praise Tencent's Hy3 model and its Apache 2.0 open licensing as meaningful progress for open AI, while negative users call the model disappointing at coding or inferior to alternatives like GLM.
Pos
76.7%
Neg
23.3%
23 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Related links
Posts from X
Most Activity
Most Activity
VIEWS34.7KBOOKMARKS62LIKES196REPLIES19
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
If these scores are real, Tencent has just become one of the leaders of open source. Hy3 is a 295B 21AB model that at least beats GLM 5.1 in blind tests. Might even warrant architecture transition (it's a basic GQA). Good job.
Florian Brand@xeophon
Tencent @TencentHunyuan dropped the non-preview version of Hy3 and changed their license from the community one (restrictive + not allowed in SK, UK, EU) to Apache 2.0!!! 🙏
2hViews 34.7KLikes 196Bookmarks 62
RETWEETS17
elie@eliebakouch
new open (apache 2.0!) model from @TencentHunyuan, Hy3 is only 295B total 21B active and competitive with MUCH bigger model on benchmarks
https://huggingface.co/tencent/Hy3
2hViews 14.1KLikes 170Bookmarks 45
elie@eliebakouch
the benchmark scores are REALLY crazy for a 300B model
elie@eliebakouch
new open (apache 2.0!) model from @TencentHunyuan, Hy3 is only 295B total 21B active and competitive with MUCH bigger model on benchmarks
https://huggingface.co/tencent/Hy3
1hViews 6.9KLikes 81Bookmarks 22
xjdr@_xjdr
Ohh this looks incredibly promising. Time to throw some internal evals at it . If it's performance is even moderately close to it's published evals, this could be a really great model
Florian Brand@xeophon
Tencent @TencentHunyuan dropped the non-preview version of Hy3 and changed their license from the community one (restrictive + not allowed in SK, UK, EU) to Apache 2.0!!! 🙏
1hViews 2.4KLikes 41Bookmarks 8
Zephyr@zephyr_z9
Pretty strong model from Tencent
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
If these scores are real, Tencent has just become one of the leaders of open source. Hy3 is a 295B 21AB model that at least beats GLM 5.1 in blind tests. Might even warrant architecture transition (it's a basic GQA). Good job.
1hViews 14.7KLikes 32Bookmarks 5
Florian Brand@xeophon
@Jaidcel @TencentHunyuan Yes! https://huggingface.co/tencent/Hy3
1hViews 441Likes 4Bookmarks 3
elie@eliebakouch
from the model card seems to not only be benchmark
> We don't think public benchmark scores tell the full story. So we ran a blind test with 270 experts from various disciplines, working on real-world workflows, and collected 312 valid comparisons. Hy3 scored 2.67/4, outperforming GLM-5.1 at 2.51/4. The advantage was clearest in frontend development, CI/CD, and data & storage.
elie@eliebakouch
new open (apache 2.0!) model from @TencentHunyuan, Hy3 is only 295B total 21B active and competitive with MUCH bigger model on benchmarks
https://huggingface.co/tencent/Hy3
2hViews 1.6KLikes 20Bookmarks 0
elie@eliebakouch
@teortaxesTex doesn't have 1M context! model without sparse attention is ok at 256k (also you can do GQA -> MSA)
1hViews 125Likes 5Bookmarks 2
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
@eliebakouch There are ways to convert GQA to MLA and therefore to DSA I mean that this model is strong but with the current baselines not economical for long sequences
1hViews 551Likes 3Bookmarks 2
Jaid@Jaidcel
@xeophon @TencentHunyuan Updated weights?
1hViews 446Likes 3Bookmarks 1
Florian Brand@xeophon
@teortaxesTex Some truly insane jumps, especially for this size
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
If these scores are real, Tencent has just become one of the leaders of open source. Hy3 is a 295B 21AB model that at least beats GLM 5.1 in blind tests. Might even warrant architecture transition (it's a basic GQA). Good job.
1hViews 4KLikes 10Bookmarks 0
elie@eliebakouch
@teortaxesTex wdym? > Might even warrant architecture transition (it's a basic GQA)
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
If these scores are real, Tencent has just become one of the leaders of open source. Hy3 is a 295B 21AB model that at least beats GLM 5.1 in blind tests. Might even warrant architecture transition (it's a basic GQA). Good job.
1hViews 929Likes 4Bookmarks 0
🜑@1casie
@xeophon @TencentHunyuan >not allowed in SK, UK, EU
1hViews 269Likes 3
dinos@din0s_
@xeophon you consistently naming and shaming non apache releases as soon as they come out has probably had a non negligible impact on the state of open models, good job
35mViews 198Likes 3
Florian Brand@xeophon
@din0s_ haha, idk how much of an impact it really has, i just don't want to deal with reading licenses :(
31mViews 169Likes 2
naeem@identity_matrix
@eliebakouch @TencentHunyuan Wow, it has improved a lot from preview.
Makes me even more excited about deepseek v4 non-preview release
1hViews 235Likes 5
何@AvimanyuRoy3
@eliebakouch @TencentHunyuan only 295B
2hViews 188
Zhu Liang@paradite_
@eliebakouch why would you take an average across different benchmarks? that doesn’t seem very scientific.
1hViews 117
elie@eliebakouch
@paradite_ This is just a quick visualization I asked Claude to make while on the plane lol, but how would you combine scores from a pool of benchmark of the same domain?
1hViews 103
elie@eliebakouch
@_xjdr Please share the result 🫡
xjdr@_xjdr
Ohh this looks incredibly promising. Time to throw some internal evals at it . If it's performance is even moderately close to it's published evals, this could be a really great model
1hViews 264Likes 3Bookmarks 0
Load more posts