@heyshrutimishra: tencent just put the weights for a 770B model on hugging face, free to download. it's called hy4, their new flagship. 4…
Summary
Tencent released open weights for HY4, a 770B multimodal flagship model with 49B active params per token and 1M token context, including instruct, FP8 and base variants. The model card also describes HY4 autonomously training a variant that beat their baseline, further narrowing the open-vs-closed gap alongside DeepSeek, Kimi and GLM releases.
View Cached Full Text
Cached at: 11/20/26, 06:05 PM
tencent just put the weights for a 770B model on hugging face, free to download.
it’s called hy4, their new flagship. 49B params active per token, 1M token context, and it reads text, images and video in one model. they released the instruct model, an fp8 version and the raw base model.
the part of the model card i keep rereading: paired with their own agent harness, hy4 optimized tencent’s pre-training setup for smaller models and then autonomously trained a variant that beat their official baseline on multiple benchmarks.
so the model helped build the next models. and they gave it away as open weights.
they’re also refreshingly honest. the card says there’s still a notable gap to the top closed models on research-level math, long-horizon coding and model self-improvement.
after deepseek v4, kimi k3 and glm 5.3, this is another chinese lab shipping frontier-scale weights in the open. the gap between open and closed keeps getting smaller, and anyone building on top of these models gets the benefit.
Similar Articles
Tencent Hy4 (flagship multimodal MoE A49B/770B) released
Tencent released open weights for Hy4, its natively multimodal flagship MoE model with 770B total/49B active parameters and 1M context, along with GitHub repo, GGUF quantizations, license, and system card.
Tencent/Hy4-preview 770B-A49B weight dropped
Tencent releases Hy4-preview, a new Mixture-of-Experts flagship AI model with 770B total parameters and 49B activated per token, featuring advanced techniques like Gated DeepSeek Sparse Attention and identity Hyper-Connections.
@AdinaYakup: Hy4's weights are now available on the hub https://huggingface.co/tencent/Hy4 - 770B/49B - 1M context - Custom license …
Tencent's Hy4 MoE flagship model weights are released on Hugging Face, featuring 770B total/49B active parameters, 1M context, multimodal text-image-video understanding, and optimized inference via DFly2 speculative decoding.
tencent/Hy4
Tencent's Hy Team released Hy4, a new-generation MoE flagship model with 770B total parameters (49B activated per token) supporting text, image, and video understanding, featuring Gated DeepSeek Sparse Attention, IndexCache, and a built-in DFly2 drafter for speculative decoding.
@cline: Tencent Hy4 Preview leads on SWE-bench Pro. 770B, 49B active, 1M context, and their biggest generational leap measured …
Tencent Hy4 Preview leads on SWE-bench Pro with 770B total parameters, 49B active, and 1M context length, marking their biggest generational leap as an open weights model competing against frontier AI.