Skip to content
Neural Daily

ModelsNews

Hy4 flash: open-weight 220B MoE with 12B active params

🇭🇺 Olvasd el magyarul →

Tencent has released Hy4 flash with open weights: a 220-billion-parameter MoE with about 12 billion active parameters per token, a 1M-token context window and image input. API pricing is $0.07 per million input tokens.

· 3 min read

Pontdiagram: nyílt agent-benchmarkok átlagpontszáma a feladatonkénti költség függvényében, a Hy4 flash a legolcsóbb, a Hy4 és a DeepSeek-V4.1-Flash felette
Image: Tencent · Screenshot · Source

What was released

Tencent announced Hy4 flash, the cheapest model in its Hy family, on November 16 (Tencent). Weights are now downloadable from Hugging Face, ModelScope and GitCode in three variants: the instruct model, an FP8-quantized instruct model and the pre-trained base model (Hugging Face). When Hy4 launched on November 4, open weights were promised for later (Tencent Hy4 adds vision and computer use, weights later).

  • 220B total parameters, including 63.5B Engram parameters (an n-gram embedding table), with about 11.3B active text parameters per token.
  • 1M-token context, text and image input, text output.
  • Hybrid attention: sparse attention plus a 512-token sliding window; the KV cache is one sixth of Hy4's.
  • A built-in 0.85B DFly2 drafter for speculative decoding that reuses the main model's KV cache.
  • API pricing: about $0.070 input and $0.281 output per million tokens. Tencent reports decoding above 150 tokens per second.

How it performs

According to Tencent's own table, Hy4 flash scores 64.6 on DeepSWE v1.1, 68.1 on OSWorld 2.1, 63.1 on APEX-Agents v1.1 and 75.8 on MMMU-Pro. Long-horizon software work is the weak spot: 19.2 on Terminal-Bench 4.0, against 44.4 for Hy4. Tencent itself recommends Hy4 for complex engineering and the flash model for research, document processing, slide decks and computer use.

Benchmarktáblázat: a Hy4 flash pontszámai és feladatonkénti költsége a Hy4, Hy4 preview, DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next, GLM-5.3-Flash és GPT-6 Luna mellett
Tencent · Screenshot · Source

Source: Tencent Hy, Hy4 flash model card (Hugging Face)

On its internal WorkIndex benchmark, Tencent puts per-task cost at 1/13 of Hy4 Preview and 1/20 of Hy4. These are vendor measurements; there are no independent results yet.

Serving and license

For production, Tencent recommends vLLM or SGLang, both with prebuilt Docker images. The sample config runs the FP8 variant with 8-way tensor parallelism and DFly speculative decoding. Transformers support currently installs from an open pull request.

The license is derived from Apache 2.0 but is not Apache 2.0. Internal and commercial use are allowed. A separate written license is required if you offer the model as a service (MaaS) and your annual revenue exceeds $50 million, or your product has more than 100 million monthly active users (Tencent Hy License). For most companies or universities self-hosting it, that is not a blocker.

Why it matters

With 12B active parameters and a small KV cache, Hy4 flash joins the cheap agent models you can realistically self-host. If you need on-prem computer use or document processing for data protection reasons, it is worth benchmarking against other open-weight MoEs, but judging by Terminal-Bench it is not the first pick for a coding agent.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles