# Hy4-flash

> Hy4 flash is our most cost-effective Hy model to date, featuring 220B total parameters (including 64B Engram parameters) with just 12B active parameters per token. It natively accepts text and image inputs, generates text, and supports a context window of up to 1M tokens.

- **Organisation:** tencent
- **Category:** Vision-Language Models
- **Parameters:** 220B
- **Context length:** 1,048,576 tokens
- **Licence:** tencent-hy-license (Check licence)
- **Released:** 2026-11-17
- **Hugging Face:** 5.6K downloads (30 days), 547 likes

## Estimated memory

| Quant | ≈ GB |
|---|---|
| Q2_K | 73.93 |
| Q3_K_M | 108.39 |
| Q4_K_M | 133.7 |
| Q5_K_M | 156.31 |
| Q6_K | 180.55 |
| Q8_0 | 231.97 |
| F16 | 433.89 |

## Runs with

- [Transformers](https://github.com/huggingface/transformers)

## About

Hy4 flash is our most cost-effective Hy model to date, featuring 220B total parameters (including 64B Engram parameters) with just 12B active parameters per token. It natively accepts text and image inputs, generates text, and supports a context window of up to 1M tokens.

Its 42-layer language backbone consists of one dense feed-forward layer and 41 MoE layers, with identity Hyper-Connections (iHC) supporting inter-layer information flow. A vision encoder (adapted from SigLIP2-So400m-Patch16-Res512) is connected to the backbone through an MLP projector. In addition to the backbone, a built-in 5-layer drafter, DFly2 (0.85B parameters), reuses the target model's KV cache and sparse indices to support long-context speculative decoding at a low, constant cost.

Compared with Hy4, Hy4 Flash introduces three notable architectural changes:

• Hybrid attention. Sparse attention (IndexPool) and sliding-window attention (SWA) are interleaved at a 1:3 ratio.
• LatentMoE. Routed experts operate in a lower-dimensional latent space. Each token activates 16 of 512 routed experts, alongside one shared expert.
• N-gram embedding (Engram). The backbone is augmented with learned bigram and trigram representations, retrieved via multi-head hash lookups.

Bits-per-byte (BPB) on internal code & agent validation, lower is better:

Few-shot pretrain benchmark evaluation:

## Alternatives

- [Qwen/Qwen3.8-Flash-Next](https://selfhost.directory/models/qwen/qwen3.8-flash-next) (180B (61.6B active), qwen-community-1.0, ≈107.96 GB at Q4): As the frontier of foundation models pushes toward ever-larger parameter counts and ever-longer context windows, the question is no longer just how much we can scale, but how efficiently we can do so.
- [MiniMaxAI/MiniMax-M3.1-Flash](https://selfhost.directory/models/minimaxai/minimax-m3.1-flash) (220B, minimax-community, ≈132.15 GB at Q4): MiniMax-M3.1-Flash is a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It supports text, image, and video inputs with a context window of up to 1M tokens.
- [zai-org/GLM-5.3-Flash](https://selfhost.directory/models/zai-org/glm-5.3-flash) (321B (18.3B active), MIT, ≈213.49 GB at Q4): 👋 Join our WeChat or Discord community. 📖 Check out the GLM-5.3-Flash blog and GLM-5 Technical report . 📍 Use GLM-5.3-Flash API services on Z.ai API Platform.
- [thinkingmachines/Inkling-Small](https://selfhost.directory/models/thinkingmachines/inkling-small) (266B (1.7B active), Apache 2.0, ≈159.48 GB at Q4): Inkling-Small is a general-purpose multimodal model that accepts text, image and audio inputs and generates text outputs. It is intended for use in English and other languages, and across multiple coding languages.
- [Qwen/Qwen3-VL-235B-A22B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-235b-a22b-instruct) (236B (22B active), Apache 2.0, ≈141.68 GB at Q4): Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.
- [CohereLabs/command-a-plus-05-2026-bf16](https://selfhost.directory/models/coherelabs/command-a-plus-05-2026-bf16) (219B (25.5B active), Apache 2.0, ≈131.19 GB at Q4): Command A+ is an open source model with 25 billion active parameters and 218B total parameters model optimized for agentic, multilingual, and reasoning-heavy tasks with a focus on enterprise performance, while also providing support for vision inputs for processing image inputs.

More: https://selfhost.directory/models/tencent/hy4-flash/alternatives

## Runs well on

- [Apple M3 Ultra (512 GB)](https://selfhost.directory/models/tencent/hy4-flash/on/apple-m3-ultra-512gb): Q8_0, est. 1.9 tok/s

## Similar size

- [Qwen/Qwen3.8-Flash-Next](https://selfhost.directory/models/qwen/qwen3.8-flash-next) (180B)
- [MiniMaxAI/MiniMax-M3.1-Flash](https://selfhost.directory/models/minimaxai/minimax-m3.1-flash) (220B)
- [Qwen/Qwen3-VL-235B-A22B-Instruct](https://selfhost.directory/models/qwen/qwen3-vl-235b-a22b-instruct) (236B)
- [thinkingmachines/Inkling-Small](https://selfhost.directory/models/thinkingmachines/inkling-small) (266B)
- [stepfun-ai/Step-3.7-Flash](https://selfhost.directory/models/stepfun-ai/step-3.7-flash) (201B)
- [FINAL-Bench/Darwin-180B-RSI-R3](https://selfhost.directory/models/final-bench/darwin-180b-rsi-r3) (180B)

## FAQ

**How much VRAM does Hy4-flash need?**
About 134 GB at Q4_K_M with an 8K context. Q8_0 needs about 232 GB and full precision about 434 GB. Longer contexts need more memory for the KV cache.

**Can Hy4-flash run on the Apple M3 Ultra (512 GB)?**
Yes. It fits at Q8_0 (about 232 GB), at an estimated 1.9 tokens/s.

**Can I use Hy4-flash commercially?**
Unclear. It is released under tencent-hy-license. Custom licence: read the model card before using it commercially. Check the licence on the model card before using it in a product.

**What is the context length of Hy4-flash?**
1,048,576 tokens (1M). The full window needs far more memory than an 8K context.

**How do I run Hy4-flash locally?**
It runs with Transformers. Each has copy-paste examples in its docs.

## Links

- huggingface: https://huggingface.co/tencent/Hy4-flash
- homepage: https://hy.tencent.ai/research/hy4-flash
- github: https://github.com/Tencent-Hunyuan/Hy4-flash

---
Source: https://selfhost.directory/models/tencent/hy4-flash