Taking control of Hy4 flash, from licensing to local deployment

Hy4 flash is now an open weights model with multimodal input, a one-million-token context window, compressed versions, and a Tencent license.

The weights for Hy4 flash are now available to download, alongside an FP8 version and a pretrained base model. This release expands deployment options for Tencent’s multimodal model, designed for everyday productivity tasks. The official repository distinguishes the instruction-tuned version from the Base model, which provides a starting point for further adaptation rather than a ready-to-use conversational assistant. Activating a Fraction of the Model for Each Token

The current model card describes approximately 220 billion total parameters, with around 12 billion activated per token. The MoE architecture selects a subset of experts at each generation step. It combines hybrid attention, experts operating in a lower-dimensional representation space, and an Engram component that retrieves learned representations of small groups of tokens. Multimodal capabilities cover text and image inputs, with text responses and a stated context window of up to one million tokens.

The team recommends the model for information retrieval, document processing, presentation creation, simple interface demos, and computer use. Instruction-tuned versions offer reasoning and direct-response modes. For production serving, the documentation recommends vLLM and SGLang; GGUF conversions are also available through the AngelSlim project. Files Remain Large for Local Inference

The AngelSlim repository contains five GGUF versions with different quantization trade-offs. File sizes range from 125.73 GiB for IQ2-AH to 156.79 GiB for Q4KM. Q3K and two SPARK-4Bit variants fall between those extremes. These names do not indicate uniform precision across all weights: the conversions retain some components at higher precision and treat experts differently. The resulting files remain substantial despite compression.

Published tests used WorkBuddy scenarios with contexts of 4,000, 16,000, and 32,000 tokens. On a Mac M5 Max with 128 GB of memory, the Q4KM and Q3K variants generate approximately 44 to 50 tokens per second, depending on context length. On the configuration labeled RTX Spark 128 GB, SPARK-4Bit-Fast reaches approximately 32 to 33 tokens per second. For code generation, reported peaks reach 70 tokens per second with Q3K on the Mac and 40 with SPARK-4Bit-Fast on the Spark machine. These measurements apply to the compressed versions and configurations described; they do not demonstrate local inference using the maximum one-million-token context.

Installation still depends on ongoing integration work. The pull request adding llama.cpp support remains open and in draft: the guide requires its specific branch, with an additional patch for the Mac configuration shown. Image analysis also requires a separate projector file. The weights are available, while this inference route still requires following the repository’s specific instructions. A Dedicated License for Large Model Services

The release uses the Tencent Hy License, Version 1.0, derived from Apache 2.0 with additional commercial terms. It permits use, modification, and redistribution under its conditions, but requires a separate written license for certain operators of remote model services, or MaaS.

This requirement applies when a MaaS operator exceeds $50 million in aggregate worldwide gross revenue over any consecutive twelve-month period, or uses the model in a product or service exceeding 100 million monthly active users. The revenue threshold applies to the operator’s overall business. The text excludes certain product-specific features without general-purpose model access from its definition of MaaS. It also provides an exception for strictly internal use when neither the model, its derivatives, nor their outputs are made available to third parties.