← all articles

what is tencent hy4? the open weights model, explained

tencent hy4 is a mixture of experts (moe) flagship model from the tencent hy team that understands text, images and video. its language backbone has 770b total parameters, with 49b activated per token, and a context length of 1m (1,048,576 tokens). as of november 2026, tencent publishes open weights for hy4, hy4-fp8 and the pre-trained hy4 base on hugging face, modelscope and gitcode.

the weights come under the tencent hy license, version 1.0, which the license text says is derived from the apache license 2.0 but is not that license.

download the ReMynd desktop app to search text you saw in a focused window, by app and date.

what did tencent release?

the hy4 model card on hugging face lists three models in its news and model links sections.

modelwhat tencent calls it
hy4instruct model
hy4-fp8fp8 quantized instruct model
hy4 basepre-trained base model, not an instruction-tuned chat model

the same text is in tencent's hy4 github repo. the card's news line gives no date. the hugging face commit history shows the repo's latest update on november 20, 2026, at 03:05 gmt.

how big is hy4?

the parameter figures in the answer above are for the language backbone. the card adds these details.

specwhat tencent lists
total parameterscounted excluding the vit, adapter and dfly2
activated per tokenthe text figure plus vision 1.23b
backbone layers78: 1 dense layer and 77 moe layers
experts per moe layer256 routed and 1 shared; each token uses the top 8 routed experts plus the shared one
drafterdfly2, a 5 layer drafter with 3.3b parameters for speculative decoding

the card says the attention module uses gated deepseek sparse attention with indexcache, inspired by deepseek and glm.

how does hy4 score?

working with tencent financial technology (fit), tencent evaluated hy4 with 213 finance professionals, including 153 experts from leading financial institutions. the card says more than 70% of these experts have over five years of industry experience, and that across 579 real-world financial tasks hy4 ranks first among open-weight models.

modelaverage score (out of 4)
hy43.04
deepseek-v4.1-flash-10172.97
kimi k3.1 pro2.68
glm-5.32.68

against hy4 preview, the card reports these changes.

  • in internal evaluations, hallucination rate cut from 15.76% to 5.45%, and speculative errors from 23.1% to 5.6%.
  • on irab, an internal benchmark from financial ai agent company rabytetech, 12.03 percentage points higher.
  • gamedev up 19.1% and gamelogic up 9.8%.

the card's known limitations section says a notable gap remains between hy4 and top proprietary models, particularly on challenging internal benchmarks for research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

what does the tencent hy license allow?

the license file says it is derived from the apache license, version 2.0, with modifications, mainly the additional commercial terms for large model as a service (maas) businesses in section 10. it also says it is not that license.

section 10 defines maas as giving third parties remote access to model inference or fine-tuning, for example via api. a maas operator must get a separate written license from tencent in either of two cases:

  1. its aggregate worldwide gross revenue exceeds us$50 million in any consecutive 12 month period, whether or not that revenue comes from hy4 or its derivative works.
  2. before using hy4 in a product or service with more than 100 million monthly active users in any calendar month.

section 10.3 says this does not apply to use solely for internal purposes, provided that neither the work, derivative works nor their outputs are made available to third parties.

how do you run hy4?

the card recommends vllm or sglang for production serving, then calling the openai-compatible api. its vllm example runs hy4-fp8 in the vllm/vllm-openai:hy4 docker image with a tensor parallel size of 8.

  • reasoning mode: reasoning_effort accepts no_think and high, and defaults to high.
  • recommended parameters: temperature 0.9 and top_p 1.0.
  • quantization: tencent provides a gguf compressed version for end to end testing purposes.

the card gives no mac instructions. for models built to fit on a laptop, see the best local llm for mac in 2026. for another new large model, see what is qwen4-max.

where does ReMynd fit?

ReMynd records your screen on a mac, runs ocr locally with apple vision, and keeps a searchable index on your computer. text comes from the focused window only, and video covers windows you have not excluded.

in our view, a model with a long context still holds only what gets sent to it. ReMynd keeps what you read on screen searchable: press control space, search in plain language, and jump to the exact moment, provided ReMynd was running then and the window was not excluded. on a mac it needs macos 14 or later on an apple silicon mac.

ReMynd also runs on windows 11 (x64 or arm64). for api prices of other models, see ai model pricing compared.

limits of this guide

  • the hy4 specs, scores and license terms here are tencent's own, from its model card, github repo and license, as read on november 21, 2026. the commit time comes from hugging face's history page. we did not run the model.
  • tencent lists the scores as its results; the finance and hallucination figures come from tencent's own evaluations.
  • we found no tencent price page we could check for the open weights, so this guide gives no price.

download the ReMynd desktop app and search what you saw in a focused window, by app and date.

common questions

what is tencent hy4? +
hy4 is a mixture of experts (moe) flagship model from the tencent hy team that understands text, images and video. its language backbone has 770b total parameters, with 49b activated per token, and a context length of 1m (1,048,576 tokens).
is hy4 open source? +
tencent publishes the hy4, hy4-fp8 and hy4 base weights on hugging face, modelscope and gitcode under the tencent hy license, version 1.0. the license says it is derived from the apache license 2.0 with added commercial terms, and is not the apache license 2.0.
can i use hy4 commercially? +
the license adds terms for model as a service (maas) businesses. a maas operator needs a separate written license from tencent if its aggregate worldwide gross revenue exceeds us$50 million in any consecutive 12 month period, whether or not that revenue comes from hy4 or its derivative works, or before using hy4 in a product with more than 100 million monthly active users in any calendar month. use solely for internal purposes is exempt, provided that neither the work, derivative works nor their outputs are made available to third parties.
can i run hy4 on a mac? +
tencent's model card recommends vllm or sglang for production serving, and its vllm example serves hy4-fp8 with a tensor parallel size of 8. it provides a gguf version for end to end testing purposes. it gives no mac instructions on the pages we read.
how good is hy4? +
tencent says hy4 ranks first among open-weight models across 579 real-world financial tasks, with an average score of 3.04 out of 4. it also says a notable gap remains between hy4 and top proprietary models particularly on challenging internal benchmarks. these are tencent's own results.