AIAny
AI Model2026
Icon for item

Hy4 flash

Handles text and image inputs and generates text, with a context window of up to 1 million tokens. Its hybrid sparse and sliding-window attention supports long-context work, while the MoE design activates about 12B parameters per token.

Introduction

A million-token context is useful only if serving it remains practical. Hy4 flash addresses that tension by combining a large, sparsely activated model with attention and decoding techniques designed to limit the cost of long-context inference. That makes it oriented toward everyday work rather than maximum capability on the hardest tasks.

Key Capabilities
  • Long-context multimodal input: It accepts text and images and supports up to 1,048,576 tokens of context, enabling workflows that involve lengthy documents or visual material.
  • Sparse computation: The model has about 220B total parameters but activates roughly 12B per token. Hybrid sparse attention and sliding-window attention, together with latent-space MoE experts, are intended to reduce inference overhead without shrinking the overall model.
  • Cost-focused decoding: Its built-in DFly2 speculative-decoding drafter reuses the target model’s KV cache and sparse indices. Tencent reports decoding speeds above 150 tokens per second with its inference optimizations; that figure depends on serving setup.
Who It Suits

A good fit for document processing, search, slide work, visual understanding, and other productivity tasks where long context and lower serving cost matter. Look elsewhere for complex, long-horizon software engineering if the highest capability ceiling is the priority: Tencent recommends Hy4 for those workloads. Reported WorkIndex results and cost comparisons come from Tencent’s internal evaluations, so they are useful signals rather than independent guarantees.

Information

Categories

More Items

Hugging Face
AI Model2026

A 4B reasoning model with refusal behavior modified using Heretic, built on a student distilled from DeepSeek-V4-Pro. Its model card reports 5 refusals in a 100-prompt test, compared with 99 for the original.

Hugging Face
AI Model2026

Generates 480p video with two-step autoregressive diffusion, reaching about 45 ms per frame on eight B200 GPUs. Distilled from Wan2.1-14B, it targets interactive generation but needs a multi-GPU setup for that speed.

Hugging Face
AI Model2026

Combines scientific data understanding with search, coding, and tool use for multi-step research. It checks intermediate results and visual feedback to revise its actions, and accepts more than ten scientific data types.