Tencent announced Hy4 on 4 November 2026 and said it "plans to release model weights and system cards in the coming days" (Tencent, 2026). As of 11 November those weights are not on Hugging Face. What is there is the earlier Hy4 preview, and it already answers the hardware question: its 4-bit GGUF is 416.97 GB, so a 512GB Mac Studio can hold it in memory. Nothing smaller than that can.
What did Tencent confirm about Hy4?
The announcement covers capabilities, price and access, not size:
- Open weights "in the coming days", with system cards, and no date.
- New in Hy4: visual understanding and computer use, for both command-line and graphical interfaces.
- Price: about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, the same as Hy4 Preview.
- Access: Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and other Tencent apps, plus OpenRouter, Hermes and OpenCode.
- Hy4 Flash, a cheaper variant, is "scheduled for release soon".
Tencent publishes its own figures, all internal: an average score of 83.18 on its WorkBuddy evaluation, and a hallucination rate down from 12.95% on Hy4 Preview to 7.2%. No independent benchmark of Hy4 exists yet.
What Tencent does not give is the parameter count of the final Hy4. Until the weights land, the preview is the only Hy4 we can size.
How big is the Hy4 preview?
tencent/Hy4-preview went up on Hugging Face on 27 August 2026 under Apache-2.0. Its model card describes a Mixture-of-Experts model with 770B total parameters, 49B activated per token, 78 layers, 256 routed experts (top 8 active) and a 1M-token context, plus a built-in 10B multi-token prediction layer for speculative decoding (tencent/Hy4-preview, 2026). The safetensors index counts 779.96B parameters in total, which is why you will see it quoted as a 780B model.
File sizes come from the Hugging Face API, read on 11 November 2026:
| Build | Repo | Size on disk | Fits 512GB? |
|---|---|---|---|
| BF16 (official) | tencent/Hy4-preview | 1,560.02 GB | No |
| FP8 (official) | tencent/Hy4-preview-FP8 | 813.8 GB | No |
| MLX 4-bit | mlx-community/Hy4-preview-4bit | 433.36 GB | Yes |
| GGUF IQ4_XS (4.33 bits per weight) | 6block/Hy4-preview-GGUF | 416.97 GB | Yes |
| GGUF Q2_K | 6block/Hy4-preview-GGUF | 289.54 GB | Yes, 2-bit |
| GGUF IQ2_XS | 6block/Hy4-preview-GGUF | 235.11 GB | Yes, 2-bit |
Get told when a better model fits your Mac Studio M5 Max 64 GB
One email when a new open-weight model is a better fit for your machine, plus the Thursday weekly.
For: Mac Studio M5 Max 64 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Can a 512GB Mac Studio run Hy4?
The preview, yes, at 4-bit. The IQ4_XS GGUF takes 416.97 GB of the 512 GB, which leaves under a fifth of the machine's memory for macOS, your apps and the KV cache. The MLX 4-bit build takes 433.36 GB and leaves less still. Both figures are file sizes; the memory used grows with the context you load, so the full 1M-token window will not fit on top of the weights.
No smaller Mac gets there at 4-bit. A 256GB machine only takes the 2-bit builds, and we do not recommend those for real work (why we never recommend 2-bit quants): the 6block model card itself measures a much larger drift from BF16 for Q2_K than for IQ4_XS.
Speed is a separate question. With 49B parameters activated per token, every token reads far more memory than mid-size MoE models do. We have no measured Hy4 speed on any Mac, so we will not print one. For what else fits this machine, see the best LLMs for a 512GB Mac Studio.
Which software runs it on a Mac?
- llama.cpp: support for the
hy_v4architecture was merged on 4 September 2026 (pull request #28127). The GGUF builds above load in a current llama.cpp. - MLX: the mlx-community build says it was converted with a branch of mlx-lm that "is not yet included in an mlx-lm release". Until it is, the MLX route means installing mlx-lm from that branch.
Whether the final Hy4 reuses the same architecture is not confirmed. If it does, these routes should carry over; if not, expect a new round of support work.
What to watch when the Hy4 weights land
Three facts decide whether Hy4 changes the picture on a Mac: its total parameter count (the memory bill), its activated count (the speed), and its license (the preview is Apache-2.0). We will update this page with the real file sizes once the repo is public.
FAQ
Are the Tencent Hy4 weights on Hugging Face?
Not as of 11 November 2026. Tencent said on 4 November that weights and system cards come "in the coming days". Only the earlier Hy4 preview, from 27 August, is published.
How much memory does Hy4 need?
For the 780B Hy4 preview, the 4-bit GGUF is 416.97 GB and the MLX 4-bit build is 433.36 GB. The official BF16 weights are 1,560.02 GB. The final Hy4 size is not published.
Can a 256GB Mac run Hy4?
Only at 2-bit (235.11 GB or 289.54 GB for the preview), which costs a lot of quality. At 4-bit, it needs a 512GB Mac Studio.
How much does the Hy4 API cost?
About US$0.90 per million input tokens and US$2.70 per million output tokens, per Tencent. Tencent also lists 50% off on TokenHub through 31 December 2026.
Sources
- Tencent, Tencent Launches Hy4: https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/
- tencent/Hy4-preview model card and file list: https://huggingface.co/tencent/Hy4-preview
- tencent/Hy4-preview-FP8: https://huggingface.co/tencent/Hy4-preview-FP8
- 6block/Hy4-preview-GGUF: https://huggingface.co/6block/Hy4-preview-GGUF
- mlx-community/Hy4-preview-4bit: https://huggingface.co/mlx-community/Hy4-preview-4bit
- llama.cpp pull request #28127: https://github.com/ggml-org/llama.cpp/pull/28127
Don't miss the next one for your Mac Studio M5 Max 64 GB
This article covers one moment. Get one email when a newer open-weight model fits your machine better, plus the Thursday weekly with what changed.
For: Mac Studio M5 Max 64 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
See how this changes your recommendation
Run the wizardHave questions? Reach out on X/Twitter